Multi-Task IndoBERT for Indonesian Capital Markets (IHSG)

Fine-tuned Multi-Task IndoBERT for empirical finance research on the IHSG. Resolves both channel register (NEWS, OPINION, OFF) and polarity (POS, NEU, NEG).

Architecture

Backbone: indobenchmark/indobert-base-p1
Head 1: channel_head (768 -> 3: OFF, NEWS, OPINION)
Head 2: pol_head (768 -> 3: NEGATIVE, NEUTRAL, POSITIVE)

Held-Out Test Set Performance (N=150)

Channel Accuracy: 90.00% (NEWS F1: 0.9515, OPINION F1: 0.7742)
Polarity Accuracy: 70.67% (Negative F1: 0.7438, Positive F1: 0.7500)
Baseline Gain: +44.67% accuracy vs. zero-shot RoBERTa (26.00%)

Quickstart Usage

import torch
from transformers import AutoModel, AutoTokenizer

repo_id = "Aoely17/indobert-ihsg-multitask"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True).eval()

text = 'IHSG melemah 0,45% ke 7.150 tertekan jual asing.'
inputs = tokenizer(text, return_tensors='pt')
with torch.no_grad():
    out = model(**inputs)
print('Channel:', out['channel_logits'].argmax(-1).item())
print('Polarity:', out['pol_logits'].argmax(-1).item())

Econometric Applications

Used to construct daily MEDIA_SENTIMENT and RETAIL_SENTIMENT indices.

Downloads last month: 26