Training KB descriptions

[Joozty]

I noticed @svlandeg trained descriptions in her Wikipedia EL script before inserting them into KnowledgeBase. What's the reason behind it? Does it bring more accurate results?

[svlandeg]

Training the descriptions is actually a crucial part of the current implementation of the entity linker. The descriptions are taken from Wikidata, and contain a lot of dense information. Basically, the EL algorithm compares the encodings of these descriptions to the sentences in which the mention occurs, to determine whether the context is similar.

You can find more information in this video given at spaCy IRL 2019, or in the more recent video tutorial.

[Joozty]

Thanks a lot. One last thing to clarify. If I understood you correctly it means that when I train my model I can remove all sentences that don't contain any entity from my corpus?

When the context in EL training is taken into account it just compares the similarity of the sentence where the entity is mentioned with vectors defined in KB, right?

[Joozty]

Oh, I see there is n_sents param to configure how many neighbours sentences take into account.

[Joozty]

I don't know where the accept button for the answer went. Maybe I don't have permission 🤔. Nonetheless, I guess this can be marked as answered. Thanks

[svlandeg]

Some of the subforums have an "open-ended" style conversation, like this "best practices" subforum. These don't have a notion of "accepted answers". Nevertheless good to hear that it's resolved!

[Joozty]

Maybe I would have one more question for you @svlandeg if you don't mind. What's the rationale behind creating custom word embeddings instead of using doc.vector property?

[svlandeg]

I don't remember all the history of the implementation details, but it would probably boil down to the same thing, yes!