Otter is a multilingual, open-type named entity recognizer. You give it a piece of text and a list of entity types in plain language -- ["person", "band", "chemical compound"] -- and it returns the character spans of the entities of those types. There is no fixed label set and no fine-tuning step: the types are part of the input.

Cross-encoder. The entity types are prepended to the text as a [LABEL] <type> ... [SEP] prefix, so a single encoder sees labels and text together and they attend to each other. More accurate than the bi-encoder, at the cost of re-encoding the text for every label set.

'Angela Merkel'           person          0.97
'Brandenburger Tor'       location        0.82
'Berlin'                  location        0.68

Each entity is a dict with text, label, start, end (character offsets into the input string) and score. Pass a list of strings to run on a batch; you then get one list of entities per input, in the same order:

predict keeps spans scoring above threshold, which defaults to config.prediction_threshold (0.5 for this checkpoint, chosen by calibrating macro-F1 across the evaluation suite). Lower it for higher recall, raise it for higher precision:

Because the label set is part of the input, the useful threshold shifts with how many types you ask for and how specific they are. If you have a few hundred annotated sentences from your own domain, re-calibrating on those is worth more than any default.

The label is read as natural language, so it carries meaning. "politician" and "person" select different spans, and a phrase like "chemical compound" works as well as a single word. Prefer the wording you would use to describe the type to a person.

The model is trained on inputs of the form

model.build_prompt(labels) returns that prefix if you want to build inputs yourself. Note that the prefix counts against max_seq_length, so very long label sets leave less room for the text.

collate_fn.py in this repository holds the training and evaluation collators. See the GitHub repository for the full training pipeline, the evaluation suite, and the data preparation scripts.

The cross-encoders are the stronger models; the bi-encoders are cheaper when one label set is applied across a large corpus.

Apache 2.0.