Train a relation extraction model with spaCy

[CarlaMPR]

Hi! :) I'm working on Relation Extraction, specifically the extraction of drug-drug interactions from text documents.
I saw this tutorial, https://www.youtube.com/watch?v=8HL-Ap5_Axo, and I'm trying to implement that in my specific case. I'm using this corpus, https://github.com/isegura/DDICorpus, consisting of a collection of XML files with the drugs and the relations between them already labelled. With this, I had two doubts and that's why I decided to send you this message.

Do I need a NER model to identify the drugs (my entities) in the first place?
In what format should I have the corpus? I saw this example, https://github.com/explosion/projects/blob/v3/tutorials/rel_component/assets/annotations.jsonl, but I couldn't understand what "token_start" and "token_end" means.
Thank You!

[polm]

Do I need a NER model to identify the drugs (my entities) in the first place?

Yes, the relex component uses existing entity annotations to find potential relations, so you need something that sets entities. Usually that would be an NER model, though it could be rule-based (an EntityRuler).

In what format should I have the corpus? I saw this example,

You need to make a .spacy file that has Docs that look like you want your output to look - they should have the right entities and the right relation data in the doc._.rel attribute. The preprocessing script in the demo project builds a file like that (it's the DocBin). In this data token_start and token_end are the token indices of the terms, as opposed to the character indices (to token_start 5 is the fifth token, etc.).

[CarlaMPR]

Thank you!!! :) Is there any example of a .spacy file? I'm not understanding very well.

[polm]

This script creates a .spacy file when it calls docbin.to_disk.

You can read more about training data in general here.

[KorneliaBastin]

Is there a way to just train the model for the selected relations or entities that are included in the training set? Or do we always have to train it for all the entities and relations? For example if I annotate documents for category x and y and relationship i and j can I then just train the model for relationship i and check the accuracy just for i? Or can I train the model for i and j and then check the accuracy for each category separately?

[polm]

Sorry for the delayed reply to this.

By default, the relation extraction model learns all the relation labels in the training set, and when predicting will consider all pairs of entities within range of each other. You can make it ignore some entity types by modifying the get_candidates function, and you can customize the scorer to only check certain relationship types or check them separately.

That said, I'm not sure why you'd train on some relationship types and then not score them at all - if your'e not interested in some it's probably best to filter the training data to remove those.