LSS_FUNC

Repo for our TOD Journal Paper . This code covers learning models using experiments in the paper, executing reranking tasks, and executing retrieval tasks.

Code structure

The codes for training models and executing reranking tasks are in bin/execute. Please check the bash files.

The codes for retrieval task and measuring the speed in retriever. In addition, the executer are in retriever/execute. Please check the bash files.

Dependencies

The code has been tested with,

pytorch==1.8.1
transformers==4.2.1
datasets==1.1.3
beir
sentence-transformers
pyserini

To use the retriever, you need in addition,

torch_scatter==2.0.6

For installtion, please execute follwing.

pip install ./

Premise

We use BEIR dataset. Please set each dataset under root_dir you set. Please see the way of prepareing dataset here.

Prepare BM25 index for reranking experiment

First, please convert files of BEIR format to pyserini format, using the following command.

$ python util/beir2pyserini.py --in_dir /path/to/BEIR/datasets/root --out_dir /path/to/each/dataset/bm25_index --sep_title

Next, please create index for BM25 using pyserini. util/index_pyserini.sh can make it. execute the bash file like following.

$ bash util/index_pyserini.sh /path/to/each/dataset/bm25_index

Note that pyserini needs anserini, a java library.

Reranking Experiment

See Readme at ./bin

Retrieval Experiment

See Readme at ./retriever

Code for Case Study

See Readme at ./analysis