bcms-bertic-ner-onnx

bcms-bertic-ner ONNX

ONNX exports of classla/bcms-bertic-ner — BERTić fine-tuned for Named Entity Recognition on BCMS (Bosnian/Croatian/Montenegrin/Serbian).

Files

Model Details

• Architecture: ElectraForTokenClassification (12 layers, 768 hidden, 12 heads, ~110M params)
• Vocabulary: 32,000 WordPiece tokens, cased (lowercase=False)
• Max sequence length: 512 tokens
• Labels (9 BIO classes):

ONNX Inputs / Outputs

Inputs (all int64 [batch_size, sequence_length]):

• input_ids
• attention_mask
• token_type_ids

Output:

• logits — float32 [batch_size, sequence_length, 9]

Usage (onnxruntime)

import onnxruntime as ort
import json
import numpy as np

with open("tokenizer.json") as f:
    tok = json.load(f)
vocab = tok["model"]["vocab"]  # {token: id}

id2label = {
    0:"B-LOC", 1:"B-MISC", 2:"B-ORG", 3:"B-PER",
    4:"I-LOC", 5:"I-MISC", 6:"I-ORG", 7:"I-PER", 8:"O"
}

session = ort.InferenceSession("model_int8.onnx", providers=["CPUExecutionProvider"])

# wp_tokenize: split text into words, greedily match longest vocab substrings,
# prefix non-initial pieces with "##", wrap with [CLS] and [SEP]

text = "Ante Babić je posjetio sjedište Europske unije u Bruxellesu."
tokens = wp_tokenize(text, vocab)  # your WordPiece tokenizer here
ids = np.array([[vocab.get(t, 1) for t in tokens]], dtype=np.int64)

logits = session.run(["logits"], {
    "input_ids": ids,
    "attention_mask": np.ones_like(ids),
    "token_type_ids": np.zeros_like(ids),
})[0]

for token, label_id in zip(tokens, logits[0].argmax(-1)):
    label = id2label[label_id]
    if label != "O":
        print(f"{token:20s} {label}")

Output:

Ante                 B-PER
Babić                I-PER
Europske             B-ORG
unije                I-ORG
Bruxellesu           B-LOC

License

Apache 2.0 — same as the original classla/bcms-bertic-ner model.

Citation

If you use this fine-tuned model, please cite the following paper:

@inproceedings{ljubesic-lauc-2021-bertic,
    title = "{BERT}i{\'c} - The Transformer Language Model for {B}osnian, {C}roatian, {M}ontenegrin and {S}erbian",
    author = "Ljube{\v{s}}i{\'c}, Nikola  and Lauc, Davor",
    booktitle = "Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing",
    month = apr,
    year = "2021",
    address = "Kiyv, Ukraine",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.bsnlp-1.5",
    pages = "37--42",
}