Dataset size: 6,927 examples.
Published splits: 5,548 train / 686 development / 693 test.
Fine-tuning seeds included in the report: [22, 42, 202].
Training: 3 epoch(s), AdamW, learning rate 2e-05, weight decay 0.01, warmup ratio 0.1.
Training batch size: 8.
Maximum sequence length: 256.
Input mode: VnCoreNLP-segmented Vietnamese.
The uploaded checkpoint is best_model from seed 202, selected by development Macro-F1.

Results

Metrics are reported as mean ± sample standard deviation over the completed seeds listed above.

Metric	Mean ± std
Test Macro-F1	0.4889 ± 0.0231
Test accuracy	0.6061 ± 0.0113
Test macro precision	0.5643 ± 0.0815
Test macro recall	0.4871 ± 0.0228
Development Macro-F1	0.4541 ± 0.0218

Per-seed results

seed	dev_macro_f1	test_macro_f1	test_accuracy
22.000000	0.458312	0.495063	0.611833
42.000000	0.430529	0.463310	0.593074
202.000000	0.473531	0.508274	0.613276

Label mapping

json

{
  "0": "Anger",
  "1": "Disgust",
  "2": "Enjoyment",
  "3": "Fear",
  "4": "Other",
  "5": "Sadness",
  "6": "Surprise"
}

Usage

python

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "BaoNhan/phobert-base-UITVSMEC"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

# Segment raw Vietnamese with VnCoreNLP before inference for this checkpoint.
text = "Đây là văn_bản tiếng_Việt đã được chuẩn_hóa ."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    probabilities = model(**inputs).logits.softmax(dim=-1)[0]

predicted_id = int(probabilities.argmax())
print(model.config.id2label[predicted_id], probabilities.tolist())

Files

aggregate_metrics.json: aggregate metrics and publishing metadata.
artifacts/per_seed_results.csv: available completed-seed results.
Other evaluation artifacts are included when present locally.

Limitations

UIT-VSMEC is small and class-imbalanced and reflects Vietnamese social-media language from a particular collection period. Emotion labels are subjective, and predictions must not be treated as psychological assessment.

Dataset citation

bibtex

@inproceedings{ho-etal-2019-emotion,
  title={Emotion Recognition for Vietnamese Social Media Text},
  author={Ho, Vong Anh and Nguyen, Duong Huynh-Cong and Nguyen, Danh Hoang and Pham, Linh Thi-Van and Nguyen, Duc-Vu and Nguyen, Kiet Van and Nguyen, Ngan Luu-Thuy},
  booktitle={Proceedings of the 16th International Conference of the Pacific Association for Computational Linguistics (PACLING 2019)},
  year={2019},
  pages={319--333},
  url={https://arxiv.org/abs/1911.09339}
}