sajida-dev/fine-tune-deberta-v3-base-end2end-absa-model

Finetuned Deberta End-to-End Aspect-based Sentiment Analysis

This model is a fine-tuned version of yangheng/deberta-v3-base-end2end-absa for Aspect-Based Sentiment Analysis (ABSA).
It performs end-to-end ABSA by jointly extracting aspect terms and their sentiments using a single token-classification head. Labels follow an IOB-with-sentiment format, for example B-ASP-Positive, I-ASP-Negative, or O for non-aspect tokens.

• Developed by: Sajida-dev
• Model type: Token Classification (Aspect-Based Sentiment Analysis)
• Language(s): English
• License: MIT
• Finetuned from: yangheng/deberta-v3-base-end2end-absa
• Library: Transformers

Model Sources

• Repository: https://huggingface.co/sajida-dev/fine-tune-deberta-v3-base-end2end-absa-model
• Base Paper: DeBERTa: Decoding-enhanced BERT with Disentangled Attention (https://arxiv.org/abs/2006.03654)
• Demo: Available via Hugging Face pipeline("token-classification")

---

Uses

Direct Use

• Extract aspect terms from text (e.g., "battery life", "screen")
• Assign sentiment polarity (Positive, Negative, Neutral) to each aspect
• Useful for product reviews, customer feedback, and opinion mining

Downstream Use

• Integration into customer service analytics
• Market research sentiment dashboards
• Fine-tuning for domain-specific ABSA tasks (restaurants, healthcare, etc.)

Out-of-Scope Use

• General sentiment classification without aspect extraction
• Non-English text
• Misuse for biased or harmful profiling

---

Bias, Risks, and Limitations

Bias

Model performance depends on the dataset used for fine-tuning. If training data is domain-specific, generalization to other domains may be limited.

Limitations

• Works best on English text
• May misclassify nuanced sentiments such as sarcasm or irony
• Aspect boundary detection may fail for complex multi-word expressions

Recommendations

Users should validate outputs on their own domain-specific data and consider further fine-tuning if needed.

How to Get Started with the Model

from transformers import pipeline

nlp = pipeline(
    "token-classification",
    model="sajida-dev/fine-tune-deberta-v3-base-end2end-absa-model"
)

text = "The battery life is amazing but the screen is dull."
results = nlp(text)
print(results)

Then, you can perform inference like this. The model will automatically find the aspects in the text and classify their sentiment.

{
   "text": "The user interface is brilliant, but the documentation is a total mess.",
   "aspect": ["user interface", "documentation"],
   "position": [[4, 19], [41, 54]],
   "sentiment": ["Positive", "Negative"],
   "probability": [[1e-05, 0.0001, 0.9998], [0.9998, 0.0001, 1e-05]],
   "confidence": [0.9997, 0.9997]
}

Training Details

Training Data

• Fine-tuned on ABSA datasets (for example, product review datasets; details to be added).
• Preprocessing includes tokenization with the DeBERTa tokenizer and IOB tagging with sentiment labels.

Training Procedure

• Optimizer: AdamW
• Loss: Cross-entropy for token classification
• Mixed precision: fp16
• Early stopping: Enabled (patience = 3)

Training Hyperparameters

Hyperparameter	Value
Learning rate	2e-5
Warmup ratio	0.1
Number of epochs	5
Train batch size (per device)	16
Eval batch size (per device)	32
Gradient accumulation steps	2
Weight decay	0.01
Label smoothing factor	0.05
Evaluation strategy	Per epoch
Save strategy	Per epoch
Metric for best model	F1
Random seed	42

---

Evaluation

Testing Data

• Held-out ABSA test split from product review datasets.

Factors & Metrics

• Accuracy
• Precision
• Recall
• F1-score

Results (Final Evaluation on Test Set)

Metric	Value
Accuracy	0.9660
Precision	0.5905
Recall	0.4684
F1-score	0.5224
Eval loss	0.1200

---

Environmental Impact

Item	Details
Hardware	GPU
Training time	~3.5 hours
Training epochs	5
Carbon footprint	Not estimated

---

Technical Specifications

• Architecture: DeBERTa-v3 base with token classification head
• Objective: Joint aspect extraction and sentiment classification
• Compute Framework: Hugging Face Transformers with PyTorch

Citation

  @misc{sajida2025absa,
    author       = {Sajida-dev},
    title        = {Fine-tuned DeBERTa-v3 Base End-to-End ABSA Model},
    year         = {2025},
    publisher    = {Hugging Face},
    howpublished = {https://huggingface.co/sajida-dev/fine-tune-deberta-v3-base-end2end-absa-model}
  }