NPRF

NPRF: A Neural Pseudo Relevance Feedback Framework for Ad-hoc Information Retrieval [pdf]

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

@inproceedings{li2018nprf,
  title={NPRF: A Neural Pseudo Relevance Feedback Framework for Ad-hoc Information Retrieval},
  author={Li, Canjia and Sun, Yingfei and He, Ben and Wang, Le and Hui, Kai and Yates, Andrew and Sun, Le and Xu, Jungang},
  booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  year={2018}
}

Requirement

Getting started

Training data preparation

To capture the top-k terms from top-n documents, one needs to extract term document frequency from index. Afterwards, you are required to generate the similarity matrix upon the query and document given the pre-trained word embedding (e.g. word2vec). Related functions can be found in preprocess/prepare_d2d.py.

Training meta data preparation

We introduce two classes for the ease of training. The class Relevance incorporates the relevance information from the baseline and qrels file. The class Result simplify the write/read operation on standard TREC result file. Other information like query idf is dumped as a pickle file.

Model training

Configure the MODEL_config.py file, then run 

python MODEL.py --fold fold_number temp_file_path

You need to run 5-fold cross valiation, which can be automatically done by running the runfold.sh script. The temp file is a temporary file to write the result of the validation set in TREC format. A training log sample on the first fold of TREC 1-3 dataset is provided for reference, see sample_log.

Evaluation

After training, the evaluation result of each fold is retained in the result path as you specify in the MODEL_config.py file. One can simply run cat *res >> merge_file to merge results from all folds. Thereafter, run the trec_eval script to evaluate your model.

Reference

Some snippets of the code follow the implementation of K-NRM, MatchZoo.