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