How does Lynn work?

In technical terms, Lynn is a boosted trees machine learning ensemble. It analyzes groups of words for similar meaning, and so can compare found clauses with a database of thousands of alternatives and variants. These have been manually screened and classified by the experts at JuriBlox and consultancy ICTRecht: very strict, very long, and so on.

This matching of texts is known as natural language processing as a field of study, which has gained a lot of notoriety in recent years. Through major innovations in hardware and software, very high reliability can be achieved in identifying texts and extracting relevant information.

In computer analysis of texts, errors are possible. Just like human analysis you will think, but a computer analyzes differently. At its core, Lynn is looking for an appropriate label to put on a clause. This can go wrong in two ways: the wrong label, or no label at all because there is too little information in the clause.

Various measures have been built against these types of errors, such as omitting recognition with poor reliability and double-checking ‘suspect’ labels using keywords. All reviewed documents are also periodically checked to improve recognition even further.

For more information you can read the technical paper by NDA Lynn creator Arnoud Engelfriet.