PhoBERT Emotion Recognition v1.1

Model Description

PhoBERT Emotion v1.1 is a fine-tuned sequence classification model based on vinai/phobert-base-v2 for 7-class emotion recognition in Vietnamese social media texts:

• Enjoyment (Thích thú / Vui vẻ)
• Sadness (Buồn bã)
• Disgust (Chán ghét / Khinh bỉ)
• Anger (Tức giận)
• Fear (Sợ hãi)
• Surprise (Ngạc nhiên)
• Other (Khác / Trung tính)

The model is trained on a curated and balanced Vietnamese social corpus (11,193 samples) with anti-overfitting techniques (Cosine Annealing scheduler, early stopping, and weight decay). In addition to standard PyTorch weights, optimized ONNX FP32 and ONNX Dynamic INT8 formats are provided for high-speed, resource-efficient CPU inference in production environments.

Model Formats & Artifacts

Intended Uses & Capabilities

• Social Listening & Monitoring: Real-time emotion sentiment tracking on comments, posts, and forum threads.
• Mental Health & Chatbot Companions: Identifying distressed, anxious, or joyful states in conversational agents.
• Customer Experience Analysis: Detecting frustration, anger, or satisfaction in feedback and review texts.

How to Use

1. Using Hugging Face pipeline

from transformers import pipeline

classifier = pipeline(
    task="text-classification",
    model="huyleit/phobert-vi-emotion-v1.1", # Or local folder path
    tokenizer="huyleit/phobert-vi-emotion-v1.1"
)

text = "Hôm nay nhận được tin vui quá trời luôn!"
prediction = classifier(text)
print(prediction)

2. Using PyTorch (transformers)

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "huyleit/phobert-vi-emotion-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

id2label = {
    0: "Enjoyment",
    1: "Sadness",
    2: "Disgust",
    3: "Anger",
    4: "Fear",
    5: "Surprise",
    6: "Other"
}

text = "Hôm nay nhận được tin trúng tuyển thực sự vui phát khóc luôn á!"
inputs = tokenizer(text, return_tensors="pt", max_length=128, truncation=True)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)

pred_id = torch.argmax(probs, dim=-1).item()
confidence = probs[0][pred_id].item()

print(f"Predicted: {id2label[pred_id]} (Confidence: {confidence:.4f})")

3. Using ONNX Runtime (Fast CPU Inference)

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

model_id = "huyleit/phobert-vi-emotion-v1.1"
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load ONNX quantized INT8 model for ultra-fast inference
session = ort.InferenceSession("onnx/phobert_emotion_int8.onnx", providers=["CPUExecutionProvider"])

id2label = {
    0: "Enjoyment",
    1: "Sadness",
    2: "Disgust",
    3: "Anger",
    4: "Fear",
    5: "Surprise",
    6: "Other"
}

text = "Thật sự rất sốc và không thể tin nổi chuyện này lại xảy ra!"
inputs = tokenizer(text, return_tensors="np", max_length=128, truncation=True)

ort_inputs = {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"]
}
logits = session.run(["logits"], ort_inputs)[0]
pred_id = int(np.argmax(logits, axis=1)[0])

print(f"Predicted: {id2label[pred_id]}")

Evaluation Results

The model was evaluated on a clean held-out test set of 1,679 samples (evaluation/emotion/data/phobert_test.json).

Classification Report (Held-out Test Set)

Runtime & Quantization Benchmark Comparison

Training Details

Training Data

• Total samples: 11,193 Vietnamese social media text samples (balanced corpus).
• Split Ratio: 70% Train (7,835 samples) / 15% Validation (1,679 samples) / 15% Held-out Test (1,679 samples).

Training Hyperparameters

• Base Architecture: vinai/phobert-base-v2 (135M parameters)
• Max Sequence Length: 128
• Batch Size: 16
• Learning Rate: 1.5e-5
• LR Scheduler: CosineAnnealingLR (Warmup steps: 200)
• Weight Decay: 0.02 (L2 Regularization)
• Precision: Mixed Precision FP16
• Optimization Metric: Validation Macro F1 (eval_f1)
• Early Stopping: Patience = 2 epochs
• Hardware: Google Colab NVIDIA Tesla T4 GPU

Label Mapping

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

Limitations and Bias

• Social Media Slang & Teencode: The model performs robustly on standard text and common social media expressions, but extreme teencode or emerging neologisms may reduce classification performance.
• Sarcasm & Irony: Sarcastic expressions where positive words carry negative emotional meaning remain inherently challenging for subword-level contextual embeddings without broader conversation history.
• Subjectivity: Boundaries between emotions like Disgust vs Anger or Other vs Sadness can be subjective across different annotators.

Citation & Acknowledgements

If you use this model or its quantized variants, please cite the following foundational works:

@inproceedings{phobert,
  title     = {{PhoBERT: Pre-trained language models for Vietnamese}},
  author    = {Dat Quoc Nguyen and Anh Tuan Nguyen},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2020},
  pages     = {1037--1042},
  year      = {2020}
}

@inproceedings{ho2020vsmec,
  title     = {Emotion Recognition for Vietnamese Social Media Text},
  author    = {Vong Anh Ho and Duong Huynh-Cong Nguyen and Danh Hoang Nguyen and Pham-Nguyen Cuong and Duc-Vu Nguyen and Kiet Van Nguyen and Ngan Luu-Thuy Nguyen},
  booktitle = {2020 17th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON)},
  year      = {2020}
}

@inproceedings{demszky2020goemotions,
  title     = {{GoEmotions: A Dataset of Fine-Grained Emotions}},
  author    = {Dorottya Demszky and Dana Movshovitz-Attias and Jeongwoo Ko and Alan Cowen and Gaurav Nemade and Sujith Ravi},
  booktitle = {58th Annual Meeting of the Association for Computational Linguistics (ACL)},
  year      = {2020}
}