Skip to content

Latest commit

 

History

History
69 lines (51 loc) · 4.14 KB

README.md

File metadata and controls

69 lines (51 loc) · 4.14 KB
Raw

HC4: HLTCOE CLIR Common-Crawl Collection

This repository contains the scripts for downloading and validating scripts for the documents. Document ids, topics, and qrel files are in resources/hc4/

Required packages for the scripts are recorded in requirements.txt.

We recommand creating a new python environment for downloading. Package versions could have some unintentional effect on decoding the documents from Common Crawl. Documents could have changed on the Common Crawl file for numerous reasons, including take down requests. When a document changes, we record them in a change log document. Please raise and issue if you have documents with mismatch hashs that are not yet recorded.

Topics and Qrels

Topics are stored in jsonl format and located in resources/hc4. The language(s) the topic is annotated for is recored in the language_with_qrels field. We provide the English topic title and description for all topics and human translation for the languages that it has qrels for. We also provide machine translation of them in all three languages for all topics. Narratives(field narratives) are all in English and has one entry for each of the languages that has qrels. Each topic also has an English report(field report) that is designed to record the prior knowledge the searcher has.

Qrels are stored in the classic TREC style located in resources/hc4/{lang}.

Download Documents

To download the documents from Common Crawl, please use the following command. If you plan to use HC4 with ir_datasets, please specify ~/.ir_datasets/hc4 as the storage or make a soft link to to the directory you wish to store the documents. The document ids and hashs are stored in resources/hc4/{lang}/ids*.jsonl.gz. Russian document ids are separated into 8 files.

python download_documents.py --storage ./data/ \
                             --zho ./resources/hc4/zho/ids.jsonl.gz \
                             --fas ./resources/hc4/fas/ids.jsonl.gz \
                             --rus ./resources/hc4/rus/ids.*.jsonl.gz \
                             --jobs 4

If you wish to only download the documents for one language, just specify the id file for the language you wish to download. In case the URLs for the Common Crawl files change in the future, the flag --cc_base_url provides the options to specify an alternative URL for the files. The current default value points to https://data.commoncrawl.org/. The full description of the arguments can be found when execute with the --help flag.

Postprocessing of the Downloaded Documents

Multiprocessing during download results in arbitrary ordering of the documents in the saved .jsonl files. To support full reproducibility, we provide script to postprocess the file to match the document order specified in the document id files. fix_document_order.py changes the ordering of the documents, validates the document hashs, and verifies all and only specified documents are in the result file. The unsorted file will be renamed as hc4_docs.jsonl.bak. You could delete the file manually. Following is a sample command.

python fix_document_order.py --hc4_file ./data/rus/hc4_docs.jsonl \
                             --id_file ./resources/hc4/rus/ids*.jsonl.gz \
                             --check_hash

If the script identifies missing files during postprocessing, please rerun the downloading script with --resume flag to get the missing documents. Some files might be missing due to temporary network failure or connection refused by the Common Crawl servers. Rerunning the downloading script usually would be able to retrieve those documents. If not, please raise issue with the document id to bring this to our attention.

Reference

If you use this collection, please kindly cite our dataset paper with the following bibtex entry.

@inproceedings{hc4,
	author = {Dawn Lawrie and James Mayfield and Douglas W. Oard and Eugene Yang},
	title = {{HC4}: A New Suite of Test Collections for Ad Hoc {CLIR}},
	booktitle = {Proceedings of the 44th European Conference on Information Retrieval (ECIR)},
	year = {2022}
}