Bi-xLSTM_Yoga

Bi-xLSTM[7:1] for Indonesian End-to-End ABSA

This repository contains a Bi-xLSTM[7:1] model pretrained on large-scale Indonesian Wikipedia data and fine-tuned for Indonesian End-to-End Aspect-Based Sentiment Analysis (E2E-ABSA).

Model Description

The model uses a bidirectional xLSTM-based architecture for contextual language modeling. It was first pretrained on Indonesian Wikipedia data using a forward and backward language modeling objective, then fine-tuned for E2E-ABSA using BIOES sentiment tagging and CRF decoding.

The final task is to extract aspect–sentiment pairs directly from Indonesian review text.

Architecture

• Model: Bi-xLSTM[7:1]
• Pretraining objective: Bidirectional contextual language modeling
• Fine-tuning task: End-to-End Aspect-Based Sentiment Analysis
• Decoder: CRF
• Labeling scheme: BIOES with sentiment labels
• Framework: PyTorch
• Language: Indonesian

Dataset

The model was pretrained using Indonesian Wikipedia data and fine-tuned on Indonesian review data for aspect-based sentiment analysis.

Intended Use

This model is intended for research and academic purposes, especially for:

• Indonesian NLP
• Sequence labeling
• Aspect-Based Sentiment Analysis
• Contextual language modeling
• Comparison between xLSTM-based models and Transformer-based models