Skip to content

Using NLP Suggestions

The editor can suggest the constituents of an act frame — actor, action, object, recipient — directly from Dutch source text, using a machine-learning model. This removes the manual first step of deciding which words play which role.

When to use it

NLP suggestions are most useful when you are starting an act on a Dutch sentence and want a head start on identifying its parts. The feature is entirely optional; you can interpret any source without it.

Dutch text only

The underlying model is trained on Dutch normative text. Suggestions on text in other languages are unreliable. The model runs on the sentences the act is anchored to, one at a time — very long sentences can exceed the model's token limit.

How to use it

  1. Open an act that is linked to source text — for example one you created by highlighting a Dutch sentence. The suggestion controls appear in the act's Roles heading only when the act is linked to source text.
  2. Optionally pick a different model in the dropdown next to Detect roles: BERTje (2022) (the default) or legal-bert-dutch-english.
  3. Click Detect roles. The editor sends each sentence the act is anchored to to the model, which labels every word as Actor, Action, Object, Recipient, or none.
  4. A FlintFiller Recommendations dialog shows each sentence with the suggested words highlighted per role. Click a highlighted fragment to Accept, Skip, or Discard it. Review every suggestion — accept the ones that are right and discard anything incorrect.
  5. Click OK to add the accepted suggestions to the interpretation as facts, anchored to their words in the text and given the matching subtype (Agent for an actor or recipient, Action, or Object). Cancel closes the dialog without changes.
  6. Place the new facts into the act's roles with Select, as you would any other fact.

If the request fails — for example because the chosen model is not available on the server — the editor shows the error Model not present on filesystem. and stops sending sentences.

Check the facts that OK adds

For each sentence, the dialog adds as many facts as you accepted, but it takes them from the start of that sentence's list of suggestions rather than from the ones you accepted. Unless you accepted the first suggestions in a sentence, compare the new facts in the Frames list with what you accepted, and delete any that are wrong.

Suggestions are a draft, not an answer

The model is a labelling aid, not an authority on the law. It does not understand claim-duty relations, preconditions, or fact subdivisions — those remain your judgement. Treat its output as a fast first draft of an act's roles that you then verify against the text.

For the technical details of the model and service, see NLP Assistance and Backend & API services.