This repository contains the dataset first described in 'Probabilistic Models of Relational Implication' (arxiv).
Holt, Xavier. "Probabilistic Models of Relational Implication." arXiv preprint arXiv:1907.12048 (2019).
Each line represents an implication rule between (partially) instantiated propositions. The implication wrote(PERSON, BOOK) -> written_by(BOOK, PERSON)
would be represented as:
PERSON, wrote, BOOK\tBOOK, written_by, PERSON\tTrue
And the negative implication rule wrote(PERSON, BOOK) -\> enjoyed_writing(PERSON, BOOK)
would be represented as:
PERSON, wrote, BOOK\tPERSON, enjoyed_writing, BOOK\tFalse
- 'Probabilistic Models of Relational Implication.pdf': The thesis describing this dataset.
- 'implications.tsv': Implications where instantiating arguments have been genericised e.g. 'ephedrine' would become 'DRUG'. Duplicate entries resulting from this genericisation have been removed. This was the representation over which the crowd workers annotated.
- 'directional_implications.tsv': Implications that only occur strictly in one direction, that is,
p(_, _) -> q(_, _)
butq(_, _) -\> p(_, _)
or vice-versa. A label ofTrue
indicates that the implication occurs from left-to-right, a label ofFalse
indicates the implication occurs from right-to-left. - 'implications_instantiated.tsv': The fully instantiated implication rules. These have mainly been included for compatibility/comparison with the original Levy et al. dataset.