Relation extraction

When to use it

• Knowledge graph construction
• Who did what to whom in news text
• Drug and dosage links in clinical notes
• Contract parties and obligations

What your file looks like

The file looks like a span file: one passage per row. The links between marked spans are made on the screen, not in the data. A file like this is enough:

id,text
1,Ahmad Shah joined Azmaval in 2026
2,Zarghona works for the university

What reviewers see

The text with the span tools, then a Link control with From, relation and To drop-downs listing the marked spans.

What a reviewer submits

A relation points from one span to another by their character ranges and carries a relation label.

{
  "answer": {
    "spans": [
      {
        "start": 0,
        "end": 10,
        "label": "PERSON"
      },
      {
        "start": 18,
        "end": 25,
        "label": "ORG"
      }
    ],
    "relations": [
      {
        "from": {
          "start": 0,
          "end": 10
        },
        "to": {
          "start": 18,
          "end": 25
        },
        "label": "works for"
      }
    ]
  }
}

What you get back

relations lists the links a majority of reviewers drew between spans that themselves reached consensus.

{
  "n": 3,
  "value": [
    {
      "start": 0,
      "end": 10,
      "label": "PERSON",
      "text": "Ahmad Shah"
    },
    {
      "start": 18,
      "end": 25,
      "label": "ORG",
      "text": "Azmaval"
    }
  ],
  "agreement": null,
  "tie": false,
  "relations": [
    {
      "from": {
        "start": 0,
        "end": 10
      },
      "to": {
        "start": 18,
        "end": 25
      },
      "label": "works for"
    }
  ]
}

Quality checks

• Relations are only as good as the spans under them, so train reviewers on spans first.
• Define direction in the guidelines: does works for point from person to organisation?
• Use three or more reviewers, because links are rarer than spans and easy to miss.

Tips

• Keep the relation list short. Three or four relation labels is a realistic target for a first run.
• If your relations always connect the same two labels, say so in the question. It reduces wrong-direction links.
• Check a few rows by hand before the full run.

Limits

Relation labels are optional on a span question. Relations are exported to Label Studio JSON with the spans they connect.