villmow / datasets_knowledge_embedding

Datasets for Knowledge Graph Completion with textual information about the entities
MIT License
268 stars 57 forks source link

More entities in textual version of WN18RR than in original version #1

Open gloryVine opened 4 years ago

gloryVine commented 4 years ago

Hey,

the dataset in the folder datasets_knowledge_embedding/WN18RR/original/ has 40943 entities, as it should be according to Dettmers et al.'s paper. Yet datasets_knowledge_embedding/WN18RR/text/ has 41105 entitites, which means it has 162 entities more than it should. Any idea why this is the case?

Filco306 commented 2 years ago

Hello @gloryVine ,

You probably do not work on this anymore, but I think this is an issue with WN18RR itself, as it only defines entities only by their offset (which yields duplicates).

I ran some statistics, and I get that 161 entities more with the textual data - in 160 cases, 2 have the same offset; while 3 have the same offset id in 1 case - yielding 161 entities. This is of course a source for false positives and noise in the data