indobert-ABSA-FnB

ABSA Model - F&B Best (Multilabel)

This directory contains a fine-tuned BERT model for Aspect-Based Sentiment Analysis (ABSA) that simultaneously analyzes all three aspects in the Food and Beverage (F&B) industry using multilabel classification.

Model Details

• Model Architecture: BERT for Sequence Classification
• Base Model: indobenchmark/indobert-base-p2
• Aspects: Food Quality, Service, Price (simultaneous detection)
• Language: Indonesian
• Task: Multilabel Sentiment Classification

Label Mapping

The model classifies text into three sentiment categories for each aspect:

Multilabel Output: The model can detect multiple aspects in a single text and assign sentiment to each aspect independently.

Evaluation Metrics

Based on the fine-tuning results, the model achieves the following performance:

Overall Performance (Epoch 4)

• Validation Loss: 0.9696
• Overall Accuracy: 90.55%
• Macro F1-Score: 90.25%

F1 Score per Aspect

Usage

This model is loaded and used by the ABSA API service. It provides simultaneous sentiment analysis for all three aspects.

from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_path = "./absa-fnb-model/model_absa_fnb_best"
model = AutoModelForSequenceClassification.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained("indobenchmark/indobert-base-p2")

Multilabel Output Example

The model can detect multiple aspects in a single review:

{
    "Food Quality": "Positive",
    "Service": "Positive", 
    "Price": "Negative"
}

Example Input: "Makanannya enak dan pelayanannya ramah, tapi harganya agak mahal."

This demonstrates the model's ability to:

• Detect Food Quality aspect with Positive sentiment
• Detect Service aspect with Positive sentiment
• Detect Price aspect with Negative sentiment
• All in a single inference pass