Skip to content

[ISWC 2023] Dense Re-Ranking with Weak Supervision for RDF Dataset Search

License

Notifications You must be signed in to change notification settings

nju-websoft/DR2

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

23 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DR2

This is the source code and data with the paper "Dense Re-Ranking with Weak Supervision for RDF Dataset Search" ISWC 2023.

Requirements

This code is based on Python 3.7+ and Pytorch 1.9+ and uses the Hugging Face Transformers library. The following is partial list of the required packages.

  • torch
  • transformers
  • datasets
  • faiss-cpu/faiss-gpu
  • pandas
  • tqdm

Compact Document Representation

We applied two RDF snippet generation methods Illusnip and PCSG to extract representative RDF triples of datasets and concatenated the results into compact documents. You can run codes from the origin repos for triple extraction and run codes in ./code/compact-document-representation for compact document creation.

Dense Re-Ranking

For the dense re-ranking models, we reused the open source code of ColBERT and DPR. You can simply download the data and modify the configuration file to train and test the dense re-ranking model. See README at ./code/dense-re-ranking-model/ColBERT and ./code/dense-re-ranking-model/DPR for details.

Coarse Tuning with Weak Supervision

Coarse-Tuning Based on Distant Supervision

The training data used for coarse-tuning based on distant supervision is from 700K+ dataset metadata crawled from open data portals. We used the title in each metadata as keyword query and masked this field in the metadata document. The keyword queries with their corresponding dataset ids are randomly split into 90% for training and 10% for validataion and are formatted as tsv files at ./data/distant-supervision/{train/valid}.tsv.

dataset_id1 query1
dataset_id2 query2
...

The collection of metadata is at ./data/distant-supervision/collection-of-metadata.rar. The metadata file is in json format as follows.

{
    [
        {
            "dataset_id": "xxx",
            "title": "xxx",
            "description": "xxx",
            "author": "xxx",
            "tags": "[\"xxx\"]",
            "url": "https://xxx.xxx"
        },
        ...
    ]
}

This set of labeled data is used for coarse-tuning based on distant supervision.

Coarse-Tuning Based on Self-Training

Source code for the document-to-query generator is at ./code/document-to-query-generator/src.

You can use following scripts at ./code/document-to-query-generator/scripts for training and prediction.

bash ./code/document-to-query-generator/scripts/train.sh
bash ./code/document-to-query-generator/scripts/predict.sh

The labeled data genrated for coarse-tuning based on self-training are at ./data/self-training. The keyword queries generated by the document-to-query generator trained for each fold with each type of document are at ./data/self-training/{metdata/data_illusnip/data_pcsg}/{metdata/data_illusnip/data_pcsg}_fold{0-4}_pred.json in json format as follows.

[
    {
        "dataset_id": xxx, 
        "prediction": "xxx"
    },
    ...
]

Fine-Tuning

We followed 5-fold cross-validation in ACORDAR, using the training and validation data from each fold for fine-tuning.

Evaluation

All the results for re-ranking experiments are at ./outputs. The result files are named as all_{snippet_generation_method}_{normal_retrieval_model}_{dense_model}_{tuning}_reranking.tsv and in TREC format. The numbers of {tuning} indicate different training strategies, with 0 indicating coarse-tuning based on distant supervision, 1 indicating coarse-tuning based on self-training, and 2 indicating normal fine-tuning. For example, all_illusnip_BM25_ColBERT_012stage_reranking.tsv means the re-ranking results of ColBERT trained by normal fine-tuning adding both coarse-tuning based on distant supervision and on self-training with Illusnip as snippet generation method and BM25 as normal retrieval model.

105	Q0	70723	1	28.631393432617188	ColBERT-BM25
105	Q0	11080	2	28.019168853759766	ColBERT-BM25
105	Q0	11768	3	27.970335006713867	ColBERT-BM25
...

For reproducing the result data for Table 1, Table 2, and Table 3 of the paper, the following script can be used to calculate the experimental metrics.

python ./code/evaluation/eval.py \
    --test_file=./outputs/all_illusnip_BM25_ColBERT_012stage_reranking.tsv

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Citation

@inproceedings{DR2,
  author       = {Qiaosheng Chen and
                  Zixian Huang and
                  Zhiyang Zhang and
                  Weiqing Luo and
                  Tengteng Lin and
                  Qing Shi and
                  Gong Cheng},
  title        = {Dense Re-Ranking with Weak Supervision for {RDF} Dataset Search},
  booktitle    = {The Semantic Web - {ISWC} 2023 - 22nd International Semantic Web Conference,
                  Athens, Greece, November 6-10, 2023, Proceedings, Part {I}},
  series       = {Lecture Notes in Computer Science},
  volume       = {14265},
  pages        = {23--40},
  publisher    = {Springer},
  year         = {2023},
  url          = {https://doi.org/10.1007/978-3-031-47240-4\_2},
  doi          = {10.1007/978-3-031-47240-4\_2}
}

About

[ISWC 2023] Dense Re-Ranking with Weak Supervision for RDF Dataset Search

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages