ag-sc/biquad

BiQuAD question answering dataset resources

BiQuAD

prerequisites

The data can be used as is, for generation (or use on custom datasets) Python >= 3.7 is required. The bundeled datalog interpreter used for answer retrieval requires node.js >= v15.5.

create a virtual environment and install required Python packages

# create environment
python3 -m venv .venv

# activate environment
source .venv/bin/activate

# install dependencies
pip install -r requirements.txt

Install node.js dependencies

cd dltools/nodeinterp/
npm install .

download ESDB

The sqlite database can be obtained via Kaggle. It should be placed in the data directory as database.sqlite.

scripts

• output_rulecount.py Makes sure the transformations in data/transformations can be loaded properly.
• kbtransform.py Contains most of the transformation logic and can be invoked directly to syntax check the rulesets. The syntax check also outputs all placeholders or "fillers" defined per ruleset.
• execdatalog.py Runs the datalog interpreter for answer retrieval, can be directly invoked to test communication.
• soccerdb.py Contains the data model and SQL queries for matches, players, and events. Can be invoked directly to show the event dump of a random match in the database.
• write_textrep.py Applies textual representations to matches and events, output is written to data/generated/reports/
• write_datalog.py Similar to the above, generates datalog representations.
• write_questions.py Generates questions in textual and datalog form, the datalog query is then tested against the knowledge base. Output is written to data/generated/questions.
• generate_unanswerable.py Generates adverserial unanswerable questions in addition to those in data/generated/questions, output is saved to data/generated/questionsfull.
• export_squad.py Generates a new squad-like 60-20-20 split from the generated content.

other data

• data/generated-content/ The fully generated train/dev/test splits.
• data/transformations/ The rulesets for transformations of match reports and questions into text and datalog.