bad dataset at first data.
"NegationRule-Animal-D2-3081" can not infer result.
Hi, thanks for opening the discussion. The PARARULE-Plus dataset is a synthetic multi-step deductive reasoning dataset which is a extention version from PARARULE (Peter Clark et al., 2020). Both PARARULE and PARARULE-Plus follow the closed-world assumption and negation as failure. Closed-world assumption means if the information is not given then it means false which is different with open-world assumption that there is an unknown situation. The first example "NegationRule-Animal-D2-3081" can be inferred using the following rules based on closed-world assumption and negation as failure. Because bald eagle is round is not given, it means bald eagle is not round is true and then we can use rule 1 to infer that bald eagle is heavy. Then we can use rule 2 to infer that bald eagle is awful based on the result from using rule 1.
Rule 1. If something is not round then it is heavy.
Rule 2. If something is heavy and not round then it is awful.
Hi, thanks for opening the discussion. The PARARULE-Plus dataset is a synthetic multi-step deductive reasoning dataset which is a extention version from PARARULE (Peter Clark et al., 2020). Both PARARULE and PARARULE-Plus follow the closed-world assumption and negation as failure. Closed-world assumption means if the information is not given then it means false which is different with open-world assumption that there is an unknown situation. The first example "NegationRule-Animal-D2-3081" can be inferred using the following rules based on closed-world assumption and negation as failure. Because bald eagle is round is not given, it means bald eagle is not round is true and then we can use rule 1 to infer that bald eagle is heavy. Then we can use rule 2 to infer that bald eagle is awful based on the result from using rule 1.
Rule 1. If something is not round then it is heavy.
Rule 2. If something is heavy and not round then it is awful.
Because bald eagle is round is not given, it means bald eagle is not round is true?
At this point, both ChatGPT and I failed. if I add rule "If something is not square then it is light", and then we can use this rule and rule 1 to infer that bald eagle is heavy and light?!
I think applying 'connection as failure' to features with multiple potential states is strange and terrible.
Hi, Thanks for your comment. This dataset is fully synthetic and considered by the closed-world assumption which is not the real-world scenario. Our focus is on the multi-step deductive reasoning within a closed-world assumption. In this dataset, we did not consider the commonsense because we want to avoid the situation that if model know the commonsense and it may derive the answer without the need of given rules and facts even though they do not have any real meaning.