CoolFace
Modelpublic

unduood/phayathaibert-absa-sports-facility

sourceHugging Faceapache-2.0updated 9mo agoView on Hugging Face
0likes24downloads
Model Card

PhayaThaiBERT for Aspect-Based Sentiment Analysis

Fine-tuned from clicknext/phayathaibert for ABSA on Thai sports facility reviews.

Performance

MetricScore
Overall F1 (Macro)0.9546
Overall F1 (Weighted)0.9870
Overall Accuracy0.9871

Per-Aspect Performance

AspectF1Accuracy
Equipment0.93910.9864
Staff0.98010.9929
Cleanliness0.95280.9850
Atmosphere0.92170.9721
Price0.95810.9893
Location0.97750.9936
Programs0.95150.9879
Amenities0.95810.9893

Aspects

  1. 1.Equipment - อุปกรณ์ออกกำลังกาย/กีฬา
  2. 2.Staff - พนักงาน/เทรนเนอร์/โค้ช
  3. 3.Cleanliness - ความสะอาด
  4. 4.Atmosphere - บรรยากาศ
  5. 5.Price - ราคา/ความคุ้มค่า
  6. 6.Location - ทำเล/การเดินทาง
  7. 7.Programs - คลาส/โปรแกรม
  8. 8.Amenities - สิ่งอำนวยความสะดวก

Sentiment Labels

  • —none - Aspect not mentioned
  • —positive - Positive sentiment
  • —negative - Negative sentiment
  • —neutral - Factual statement without sentiment

Usage

See inference.py in this repository for usage example.

Training Details

  • —Base Model: clicknext/phayathaibert
  • —Learning Rate: 2e-05
  • —Batch Size: 16
  • —Weight Decay: 0.01
  • —Label Smoothing: 0.1
  • —Dropout: 0.3

Anti-Overfitting Techniques

  1. 1.Early Stopping (patience=3)
  2. 2.Weight Decay (L2 regularization)
  3. 3.Dropout
  4. 4.Label Smoothing
  5. 5.Gradient Clipping
  6. 6.Cosine LR Schedule with Warmup
  7. 7.Multi-label Stratified Split (Sechidis et al., 2011)

License

Apache 2.0