Natural Language Inference problems specialized for spatial semantics & reasoning.
The NLI problem patterns and automatically generated NLI sample problems are under dataset/
.
The dataset of NLI sample problems is generated from the patterns (using the mini world with selection restrictions and a toy grammar) with the following command:
python/generate_nli.py --sr config/selection_restriction.yaml --gr python/base_grammar.fcfg dataset/problem_patterns.xml -v --out dataset/160x200.json --out-max-seed 200
Note that changes to the selection restriction file can result in a different sample dataset as the fixed number of problems (e.g., 200) per pattern are randomly drawn from the all possible samples per pattern.
Abzianidze, L., Zwarts, J., Winter, Y. (2023). SpaceNLI: Evaluating the Consistency of Predicting Inferences In Space. In Proceedings of the 4th Natural Logic Meets Machine Learning Workshop (NALOMA IV).