GDESA: Greeedy Diversity Encoder with Self-Attention for Search Result Diversification

Notes

This repo provides code, retrieval results, and trained models for our following papers:

• GDESA: Greedy Diversity Encoder with Self-Attention for Search Results Diversification.
• The previous version is Diversifying Search Results using Self-Attention Network

Instructions

Trained models and baseline runs are listed in models/ and baselines/.

Data Preparation

GDESA is based on the same preprocessed data as DSSA. You can download and decompress data_cv.tar.gz from the repo of DSSA. Notice that the data folder in DSSA is also required.

Dependencies

See requirements.txt for more details. The requirements of GDESA is almost the same with DSSA, while tensorflow is replaced with torch and torchtext.

Reproduce Experiments

Run infer_reproduce.py to reproduce the 5-fold cross validation based on 5 different models. The ranking results will be written into result.json

The list-pairwise training samples should be deployed as compressed pickles, use data_pickle.py to do this. When all the pickles are generated, run train.py to train the model.