I'm a Rails developer with a lot of web experience, but none (still) in AI.

I'm working in a web text editor that judges use to writing their sentences.

The goal is to start to use AI to help the judge rule the case, either based on his own previous rulings, either based on his colleagues rulings.

The judge would provide the text for the plaintiffs and defendants petitions, and based on these two inputs the system would recommend previous rulings that apply for the case.

I already have a considerable dataset of judges rulings inside the database, and they can be easily attached to the plaintiffs and defendants petitions for training (so this plaintiff petition + this defendant petition = this ruling).

This is specially challenging because the complaints can contain different subjects combined into the same petition; but the fact is that many offices use the same standardized petitions, as the defendants do as well, so I think the system can have a great chance of prediction success.

What algorithms or strategies should I start studying to tackle this problem?

Any similar articles, white papers or repositories that could help in my goal?

  • 1
    $\begingroup$ I hope this isn't actually going to be used in production... AIs are actually used in the courthouse, and (unsurprisingly) they turn out to be extremely racist or otherwise biased, with disastrous results. $\endgroup$
    – forest
    Mar 22 '19 at 3:40
  • $\begingroup$ @forest the project is very proof of concept, and it's not planned to be used in criminal cases. Only in civil trials, like bank agreements disputes, with a lot of thesis repetition and very little (if not none) evidences / testimonials. $\endgroup$
    – sandre89
    Mar 22 '19 at 15:33

Genuine success in this area would be beyond the state-of-the-art in research, since it likely requires analogising from relational knowledge extracted from text. In recent years, techniques for working with natural language have tended to be statistical, and are therefore somewhat deficient in this respect. You could look at 'bag of words'/latent semantic analysis approaches, but they are likely to generate many false positives unless a lot of ad hoc conditions are added manually. More recent work on 'treenets' (paper here) is more structurally informed, but is still a relatively new area.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.