phobert-product-ner

This model is a fine-tuned version of vinai/phobert-base-v2 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.0037
• Precision: 0.5
• Recall: 0.7
• F1: 0.5833
• Accuracy: 0.8077

Model description

Input

Important: The input text must already be word-segmented.

For example:

Tôi muốn mua sữa_chua vinamilk

not:

Tôi muốn mua sữa chua vinamilk

The model does not perform Vietnamese word segmentation itself.

Usage

The easiest way to use the model is with the Hugging Face pipeline API.

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="hoaan/phobert-product-ner",
    aggregation_strategy="simple"
)

text = "tôi muốn mua sữa_chua vinamilk"

result = ner(text)

print(result)

The model will return a list of detected entities:

[
    {
        'entity': 'B-PRODUCT_NAME', 
        'score': np.float32(0.4434441), 
        'index': 4, 
        'word': 'sữa_chua', 
        'start': None, 
        'end': None
    }, 
    {
        'entity': 'B-PRODUCT_NAME', 
        'score': np.float32(0.1281703), 
        'index': 5, 
        'word': 'vin@@', 
        'start': None, 
        'end': None
    }, 
    {
        'entity': 'B-PRODUCT_NAME', 
        'score': np.float32(0.13579331), 
        'index': 6, 
        'word': 'amil@@', 
        'start': None, 
        'end': None
    }
]

The exact scores depend on the input and model prediction.

Entity Labels

The model uses BIO-style labels internally, for example:

B-PRODUCT
I-PRODUCT
B-CATEGORY_NAME
I-CATEGORY_NAME
O

When using:

aggregation_strategy="simple"

Hugging Face automatically aggregates the token-level predictions into entity-level results.

For example:

B-PRODUCT I-PRODUCT

is returned as a single:

PRODUCT

entity.

Example

from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="hoaan/phobert-product-ner",
    aggregation_strategy="simple"
)

text = "tôi đang tìm sữa_chua vinamilk"

entities = ner(text)

for entity in entities:
    print(
        entity["entity_group"],
        entity["word"],
        entity["score"]
    )

Using the Model Directly

If you need lower-level access, you can load the tokenizer and model directly:

from transformers import (
    AutoTokenizer,
    AutoModelForTokenClassification
)

model_id = "hoaan/phobert-product-ner"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForTokenClassification.from_pretrained(
    model_id
)

For most users, however, the pipeline API is recommended.

Notes

• The model is designed for Vietnamese text.
• Input should be Vietnamese text that has already been word-segmented.
• The model performs token classification / NER; it does not perform word segmentation.
• aggregation_strategy="simple" is recommended when you want entity-level results instead of individual token predictions.

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 50

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	1.0	7	1.9580	0.0	0.0	0.0	0.6923
No log	2.0	14	1.7075	0.0	0.0	0.0	0.6923
No log	3.0	21	1.6631	0.0	0.0	0.0	0.6923
No log	4.0	28	1.6466	0.0	0.0	0.0	0.6923
No log	5.0	35	1.5872	0.0	0.0	0.0	0.6923
No log	6.0	42	1.4829	0.0	0.0	0.0	0.6923
No log	7.0	49	1.4084	0.0	0.0	0.0	0.6923
No log	8.0	56	1.3170	0.125	0.1	0.1111	0.7115
No log	9.0	63	1.3044	0.25	0.3	0.2727	0.75
No log	10.0	70	1.2475	0.25	0.3	0.2727	0.75
No log	11.0	77	1.2179	0.2143	0.3	0.25	0.75
No log	12.0	84	1.1870	0.2143	0.3	0.25	0.75
No log	13.0	91	1.1580	0.2857	0.4	0.3333	0.75
No log	14.0	98	1.1199	0.3571	0.5	0.4167	0.7692
No log	15.0	105	1.0870	0.3571	0.5	0.4167	0.7692
No log	16.0	112	1.1182	0.4286	0.6	0.5	0.7692
No log	17.0	119	1.0651	0.4286	0.6	0.5	0.7885
No log	18.0	126	1.0378	0.4286	0.6	0.5	0.7692
No log	19.0	133	1.0537	0.4286	0.6	0.5	0.7885
No log	20.0	140	1.0368	0.4286	0.6	0.5	0.7885
No log	21.0	147	1.0037	0.5	0.7	0.5833	0.8077
No log	22.0	154	1.0368	0.4667	0.7	0.56	0.7885
No log	23.0	161	1.0143	0.5	0.7	0.5833	0.8077
No log	24.0	168	1.0121	0.4667	0.7	0.56	0.7885
No log	25.0	175	0.9948	0.5	0.7	0.5833	0.8077
No log	26.0	182	0.9946	0.5	0.7	0.5833	0.8077
No log	27.0	189	0.9840	0.5	0.7	0.5833	0.8077
No log	28.0	196	0.9727	0.5	0.7	0.5833	0.8077
No log	29.0	203	0.9724	0.5	0.7	0.5833	0.8077
No log	30.0	210	0.9705	0.5	0.7	0.5833	0.8077
No log	31.0	217	0.9850	0.4667	0.7	0.56	0.7885
No log	32.0	224	0.9834	0.4667	0.7	0.56	0.7885
No log	33.0	231	0.9795	0.4667	0.7	0.56	0.7885
No log	34.0	238	0.9681	0.4667	0.7	0.56	0.7885
No log	35.0	245	0.9727	0.4375	0.7	0.5385	0.7885
No log	36.0	252	0.9733	0.4667	0.7	0.56	0.7885
No log	37.0	259	0.9714	0.4667	0.7	0.56	0.7885
No log	38.0	266	0.9659	0.4667	0.7	0.56	0.7885
No log	39.0	273	0.9725	0.4667	0.7	0.56	0.7885
No log	40.0	280	0.9794	0.4667	0.7	0.56	0.7885
No log	41.0	287	0.9656	0.4667	0.7	0.56	0.7885
No log	42.0	294	0.9524	0.5	0.7	0.5833	0.8077
No log	43.0	301	0.9548	0.5	0.7	0.5833	0.8077
No log	44.0	308	0.9585	0.4667	0.7	0.56	0.7885
No log	45.0	315	0.9587	0.4667	0.7	0.56	0.7885
No log	46.0	322	0.9608	0.4667	0.7	0.56	0.7885
No log	47.0	329	0.9567	0.4667	0.7	0.56	0.7885
No log	48.0	336	0.9533	0.4667	0.7	0.56	0.7885
No log	49.0	343	0.9520	0.4667	0.7	0.56	0.7885
No log	50.0	350	0.9520	0.4667	0.7	0.56	0.7885

Framework versions

• Transformers 5.16.1
• Pytorch 2.11.0+cu128
• Datasets 4.0.0
• Tokenizers 0.23.1