sdd-sentiment-kepolisian

Model Card: Sentiment Kepolisian

Model Overview

Model Name: sdd-sentiment-kepolisian Base Model: indobenchmark/indobert-base-p2 Task: Sentiment classification (4-class: negatif, netral, positif, irrelevant) Language: Indonesian Domain: Police-related news articles

---

Model Description

Fine-tuned IndoBERT for sentiment analysis of Indonesian police-related news articles. Trained on full article text (title + fulltext, untruncated) rather than a short extracted excerpt — full context proved essential for distinguishing sentiment classes in this domain (see optimization report for the ablation that established this).

Classes:

• negatif (Negative)
• netral (Neutral)
• positif (Positive)
• irrelevant (Not relevant to policing sentiment)

---

Performance Metrics

Target: Macro F1 >= 0.80 and Recall (negatif) >= 0.85 — both PASS.

Per-Class Performance (test set, n=200/class)

netral is the weakest class across every experiment run for this model, and remains the main area for future improvement (see optimization report).

---

Usage

Load Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "AzrilFahmiardi/sdd-sentiment-kepolisian"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

Inference

def analyze_sentiment(title: str, fulltext: str) -> dict:
    text = f"{title} {fulltext}"
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits

    probabilities = torch.softmax(logits, dim=-1)[0].cpu()
    predicted_class = logits.argmax(-1).item()
    predicted_label = model.config.id2label[predicted_class]
    confidence = probabilities[predicted_class].item()

    return {
        "sentiment": predicted_label,
        "confidence": confidence
    }

# Example
title = "Polisi Tangkap Pelaku Curanmor di Jakarta Selatan"
fulltext = "Kepolisian Resor Jakarta Selatan berhasil menangkap seorang pelaku pencurian kendaraan bermotor yang telah beraksi di beberapa lokasi..."
result = analyze_sentiment(title, fulltext)
print(f"Sentiment: {result['sentiment']} ({result['confidence']:.2%})")

Output Format

{
  "sentiment": "negatif",
  "confidence": 0.8734
}

---

Input/Output

---

Preprocessing Note

This model expects title + full article text, untruncated, as input. Do not manually extract or truncate the article to a short excerpt before tokenizing — that preprocessing was tested and found to significantly degrade performance (Macro F1 dropped from ~0.86 to ~0.78, mainly on the netral class). The tokenizer's own 512-token truncation already handles arbitrarily long articles safely; no crash risk regardless of input length.