GLiNER2

GLiNER2 is a small AI model built to pinpoint specific pieces of information inside documents. Finding names, dates, or clauses in a contract is called information extraction, and GLiNER2 was designed to handle this task cheaply while still offering some flexibility.

The way GLiNER2 finds information is similar to cutting out every possible section of a long book in advance and checking each one against the answer. It generates a span for every possible starting position and length in a sentence, then checks one by one which of these spans matches the entity being searched for. The problem was that the number of candidate spans exploded as length increased. So spans were usually capped at around twelve words, which meant longer phrases, like a forty-word liability clause, fell outside what the model could ever find in the first place.

This limitation was addressed differently in the version that followed. Instead of enumerating whole spans, the newer approach shifted to marking only the start and end points, making it possible to find long phrases at the same cost as short names.

The article mentions GLiNER2 while comparing it to its successor, explaining that "the earlier version, GLiNER2, enumerated candidate spans one by one when searching for entities." Despite the similar names, GLiNER2 and its successor are separate models that use fundamentally different approaches to finding spans.