This is a Keras (TensorFlow) implementation of the model described in:

Kai Hui, Andrew Yates, Klaus Berberich, Gerard de Melo. PACRR: A Position-Aware Neural IR Model for Relevance Matching. In EMNLP, 2017.

Kai Hui, Andrew Yates, Klaus Berberich, Gerard de Melo. Co-PACRR: A Context-Aware Neural IR Model for Ad-hoc Retrieval. In WSDM, 2018.

Code author: Kai Hui and Andrew Yates

Pull requests and issues: @khui @andrewyates

The (RE-)PACRR models the interaction between a query-document pair, evaluating the relevance of the document for the query. The input is the similarity matrix between a query and a document, and the output is a scaler. The full pipeline of the model will be published soon.

First install Anaconda (Recommended).

• Anaconda (instructions)

Thereafter install tensorflow and keras on Anaconda.

• TensorFlow (instructions)
• Keras (instructions)
• numpy and scipy (instructions)
• sacred (instruction)

The preprocessing is to generate a simmilarity matrix for individual query-document pairs.

There are two phases:

1. Extract the text from warc;

2. Prepare the word embedding for individual terms in the query and docs, by using a pre-trained word2vec corpus.One needs to make up the missing vectors for the terms that are not included in the pre-trained word2vec. In the preprocess/wordembedding directory, the train_w2v.py makes up the missing ones by keep training the word2vec meanwhile fixing the vectors that are presented already;

3. the computation of the similarity matrices for individual query-doc pairs (not included in the code)

At this moment, we include the pre-computed simmilarity matrices in the followings. One could download and unpack the similarity matrices for clueweb as described in PACRR and RE-PACRR.

run the following:

       tar xvf simmat.tar.gz

Configure the $parentdir in *.sh as the root directory for all outputs.

Configure the sim_dir in utils/config.py, holding the similarity matrices.

python -m train_model with expname=$expname train_years=$train_years {param_name=param_val}

or use the script

bash bin/train_model.sh

Configure different parameters in train.sh or utils/config.py

python -m pred_per_epoch with expname=$expname train_years=$train_years test_year=$test_year {param_name=param_val}

or use the script

bash bin/pred_per_epoch.sh

Configure different parameters in pred_per_epoch.sh or utils/config.py

Evaluate the prediction over the three benchmarks as described in our RE-PACRR paper. Note that for Rerank-ALL benchmark one needs to dump all trec-runs and their corresponding evaluation results under data/trec_runs and data/eval_trec_runs respectively.

The evaluation code strictly corresponds to the exp reported in PACRR and REPACRR papers, but one could easily develop evaluation pipeline for different dataset or even different benchmarks based on the code. For example, the usage of the six years' Web Track for training/validation/test is hard coded in the train_test_years in utils/config.py, and one needs to edit it when fewer years are used.

python -m evals.docpairs with expname=$expname train_years=$train_years {param_name=param_val}

python -m evals.rerank with expname=$expname train_years=$train_years {param_name=param_val}

or use the script

bash bin/evals.sh

Configure different parameters in evals.sh or utils/config.py

If you use the code, please cite the following papers:

@inproceedings{hui2017pacrr,
  title={PACRR: A Position-Aware Neural IR Model for Relevance Matching},
  author={Hui, Kai and Yates, Andrew and Berberich, Klaus and de Melo, Gerard},
  booktitle={Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
  pages={1060--1069},
  year={2017}
}

@inproceedings{hui2018co,
  title={Co-PACRR: A Context-Aware Neural IR Model for Ad-hoc Retrieval},
  author={Hui, Kai and Yates, Andrew and Berberich, Klaus and de Melo, Gerard},
  booktitle={Proceedings of Web Search and Data Mining 2018},
  year={2018},
  location={Los Angeles, CA USA}
}