Labels decoder

[Ingvarstep]

Implement a joint encoder–decoder architecture for NER.
The encoder models spans and conditions the label generation performed by the decoder. Additionally, prompt-based decoding is supported, enabling label generation based on prompt anchors.

---

Architecture

---

Usage

from gliner import GLiNER
model = GLiNER.from_pretrained('model-path')

text = "Apple was founded as Apple Computer Company on April 1, 1976, by Steve Wozniak, Steve Jobs (1955–2011) and Ronald Wayne to develop and sell Wozniak's Apple I personal computer."

labels = ["person", "other"]

model.run(texts, labels, threshold=0.3, num_gen_sequences=1)

Example output:

[
  [
    {
      "start": 21,
      "end": 26,
      "text": "Apple",
      "label": "other",
      "score": 0.6795641779899597,
      "generated labels": ["Organization"]
    },
    {
      "start": 47,
      "end": 60,
      "text": "April 1, 1976",
      "label": "other",
      "score": 0.44296327233314514,
      "generated labels": ["Date"]
    },
    {
      "start": 65,
      "end": 78,
      "text": "Steve Wozniak",
      "label": "person",
      "score": 0.9934439659118652,
      "generated labels": ["Person"]
    },
    {
      "start": 80,
      "end": 90,
      "text": "Steve Jobs",
      "label": "person",
      "score": 0.9725918769836426,
      "generated labels": ["Person"]
    },
    {
      "start": 107,
      "end": 119,
      "text": "Ronald Wayne",
      "label": "person",
      "score": 0.9964536428451538,
      "generated labels": ["Person"]
    }
  ]
]

---

You can constrain the set of possible labels generated by the decoder:

model.run(
    texts, labels,
    threshold=0.3,
    num_gen_sequences=1,
    gen_constraints=[
        "organization type", "city", "organization",
        "technology", "date", "person"
    ]
)

---

Two implementations of the label trie are available. For the more efficient C++-based version, install Cython:

pip install cython

This improves memory efficiency and computational performance when working with millions of labels.