Wiki Entity Summarization Pre-processing

Overview

This project focuses on the pre-processing steps required for the Wiki Entity Summarization (Wiki ES) project. It involves building the necessary databases and loading data from various sources to prepare for the entity summarization tasks.

Server Specifications

For the pre-processing steps, we used an r5a.4xlarge instance on AWS with the following specifications:

• vCpu: 16 (AMD EPYC 7571, 16 MiB cache, 2.5 GHz)
• Memory: 128 GB (DDR4, 2667 MT/s)
• Storage: 500 GB (EBS, 2880 Max Bandwidth)

Getting Started

To get started with the pre-processing, follow these steps:

1. Build the wikimapper database:

pip install wikimapper

If you would like to download the latest version, run the following:

EN_WIKI_REDIRECT_AND_PAGES_PATH={your_files_path}
wikimapper download enwiki-latest --dir $EN_WIKI_REDIRECT_AND_PAGES_PATH

After having enwiki-{VERSION}-page.sql.gz, enwiki-{VERSION}-redirect.sql.gz, and enwiki-{VERSION}-page_props.sql.gz loaded under your data directory, run the following commands:

VERSION={VERSION}
EN_WIKI_REDIRECT_AND_PAGES_PATH={your_files_path}
INDEX_DB_PATH="`pwd`/data/index_enwiki-$VERSION.db"
wikimapper create enwiki-$VERSION --dumpdir $EN_WIKI_REDIRECT_AND_PAGES_PATH --target $INDEX_DB_PATH

2. Load the created database into the Postgres database: read pgloader's document for the installation

./config-files-generator.sh
source .env
cat <<EOT > sqlite-to-page-migration.load
load database
    from $INDEX_DB_PATH
    into postgresql://$DB_USER:$DB_PASS@$DB_HOST:$DB_PORT/$DB_NAME
with include drop, create tables, create indexes, reset sequences
;
EOT

pgloader ./sqlite-to-page-migration.load

3. Correct missing data: After running the experiments, some issues were encountered with the wikimapper library. To correct the missing data, run the following script:

python3 missing_data_correction.py

Data Sources

The pre-processing steps involve loading data from the following sources:

• Wikidata, wikidatawiki latest version: First, download the latest version of the Wikidata dump. With the dump, you can run the following command to load the metadata of the Wikidata dataset into the Postgres database, and the relationships between the entities into the Neo4j database. This module is called Wikidata Graph Builder (wdgp).

  docker-compose up wdgp

• Wikipedia, enwiki lastest version: The Wikipedia pages are used to extract the abstract and infobox of the corresponding Wikidata entity. The abstract and infobox are then used to annotate the summary in Wikidata. To provide such information, you need to load the latest version of the Wikipedia dump into the Postgres database. This module is called Wikipedia Page Extractor ( wppe).

  docker-compose up wppe

Summary Annotation

When both datasets are loaded into the databases, we start processing all the available pages in the Wikipedia dataset to extract the abstract and infobox of the corresponding Wikidata entity. Later, these pages are marked from the extracted data, and the edges containing the marked pages are marked as candidates. Since Wikidata is a heterogeneous graph with multiple types of edges, we need to pick the most relevant edge as a summary between two entities for the summarization task. This module is called Wiki Summary Annotator (wsa), and we use DistilBERT to filter the most relevant edge.

docker-compose up wsa

Conclusion

By running the above commands, you will have the necessary databases and data loaded to start the Wiki Entity Summarization project. The next steps involve providing a set of seed nodes based on your preference along with other configuration parameters to get a fully customized Entity Summarization Dataset.

Citation

If you use this project in your research, please cite the following paper:

@misc{javadi2024wiki,
    title = {Wiki Entity Summarization Benchmark},
    author = {Saeedeh Javadi and Atefeh Moradan and Mohammad Sorkhpar and Klim Zaporojets and Davide Mottin and Ira Assent},
    year = {2024},
    eprint = {2406.08435},
    archivePrefix = {arXiv},
    primaryClass = {cs.IR}
}

License

This project is licensed under the CC BY 4.0 License. See the LICENSE file for details.