Joint Extraction

Joint extraction means finding entities in a document (names, dates, organizations, and so on) and finding the relationships between them aren't handled as two separate tasks. They're processed together in one go.

Imagine one person finds people in a photo, another finds objects, and then a third person later matches them up, saying 'this person is holding this object.' Compare that to a single person looking at the photo and immediately writing down 'this person, this object, holding' all at once. The end result might look similar, but the process is completely different. Joint extraction is closer to the second case. Because entity recognition and relationship linking happen within the same computation at the same time, there's less room for the two results to end up inconsistent with each other.

This is especially useful when you set rules in advance, like 'this relationship can only exist between these two specific types of entities.' During joint extraction, the system can pick out only the combinations that satisfy those rules, producing a complete picture (something like a table showing who is related to what, and how).

The article explains it this way: "There's also a joint extraction feature, where once you declare entity types, relationships, and structural rules like unique_head=True, beam search assembles a graph that never violates the schema." Here, joint extraction means that name recognition and relationship linking happen together within a single process inside one model, not as separate steps that get merged afterward. It's not simply about combining multiple results later.

You can try something similar to joint extraction with a chatbot.

1. Prepare a short news article or introduction text.
2. Ask the chatbot: "Find all the people's names and organization names in the text below, and also put together a table showing how each person relates to which organization (for example: works at, belongs to, represents)."
3. Notice that the result doesn't come as a separate list of entities and a separate list of relationships. Instead, they're connected together within a single table.