Missing entity pairs

[hvthaibk]

Hello,

Thanks for sharing the code and datasets. They are very helpful.
I've discovered that there are some pairs in the original ground-truth are missing in your training, validation, and testing pairs.

For example, both 19156 and 34323 do not appear in

• ContEA/datasets/FR-EN/base/train_links
• ContEA/datasets/FR-EN/base/valid_links
• ContEA/datasets/FR-EN/base/test_links

While they do appear in ContEA/datasets/FR-EN/ent _dict file as

• 19156: http://fr.dbpedia.org/resource/Université_Lille_I
• 34323: http://dbpedia.org/resource/Lille_University_of_Science_and_Technology

And the pair ("http://fr.dbpedia.org/resource/Université_Lille_I", "http://dbpedia.org/resource/Lille_University_of_Science_and_Technology") does exist in the original ground-truth.

There are 329 such pairs for FR-EN dataset.
Could you please double check? Thanks!

[audreycs]

Hi, thank you for providing your feedback. I have reviewed the issue that you raised and you are correct.

The reason for the missing ground-truth pairs is that when we constructed datasets, we randomly droped some ground-truth pairs to create a mass of dangling entities (entities without alignment) in both base KGs. We built the base KGs founded on DBP15k, where all entities have an equivalent. However, as more triples are added in the later time, a large number of dangling entities are introduced. We noticed the inconsistencies between the base KGs and the later KGs in terms of dangling entities, so we intentionally create some dangling entities to make the base KGs more realistic. We did not include this detail in the paper, but we appreciate you bringing this discover to our attention.

You can add those missing ground-truth pairs into specific files (with a ratio of 2:1:7 for the train, valid and test pairs). We expect the experimental results will align with what we reported in the paper.

[hvthaibk]

Thanks for your quick reply and your effort to mine more entity pairs from DBP15K.

We are aware that 15K pairs in DBP15K is only a partial ground-truth. And a partial ground-truth makes performance benchmarking difficult.

Do you think combining these 329 dropped pairs with the 18638 pairs in ContEA/datasets/FR-EN/base (train + valid + test) would cover all the possible entity pairs in the original dataset?