AvinabhDutta-Dev/assamese-sentiment-xlmr

Assamese Sentiment Analysis — Fine-Tuned XLM-R

This repository contains a fine-tuned XLM-RoBERTa (xlm-roberta-base) transformer model optimized for binary sentiment analysis on Assamese textual reviews. It effectively maps cross-lingual features into a single-label distribution (Positive and Negative).

---

Model Summary

• Model Architecture: XLM-RoBERTa (xlm-roberta-base)
• Task: Binary Sentiment Classification
• Language: Assamese (as)
• Developer: Avinabh Dutta
• Primary Framework: PyTorch & Hugging Face Transformers
• Top Performance: 94.88% Accuracy | 0.9452 Macro F1

---

Architecture Configuration

The implementation leverages the standard base architecture rather than the large variant:

• Hidden Size: 768
• Attention Heads: 12
• Hidden Layers: 12
• Vocabulary Size: 250,002 tokens

---

Performance & Evaluation Metrics

The fine-tuned XLM-R model achieved a robust second-place finish in the project benchmarking suite, proving highly effective at cross-lingual representation:

Metric	Score	Percentage
Accuracy	0.948832	94.88%
Macro F1-Score	0.945181	94.52%
Macro Precision	0.943848	94.38%
Macro Recall	0.946586	94.66%
ROC-AUC	0.985455	98.55%
Cohen's Kappa	0.890366	89.04%

---

Dataset Overview

The dataset consists of localized social media commentaries, media assessments, and audience reviews covering drama (নাটক) and cultural content in Assamese.

• Feature Representation: Focuses on extracting linguistic subtext from Assamese textual tokens, tracking features such as character length, word density, punctuation patterns, and emoji occurrences.
• Target Classes: Binary Classification (0: Negative, 1: Positive).

---

How to Use

You can easily load and run inference with this model using the Hugging Face transformers library:

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

# Load model and tokenizer
model_name = "AvinabhDutta-Dev/assamese-sentiment-xlmr"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Create pipeline
classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer)

# Example text (Assamese)
sample_text = "চলচ্চিত্ৰখনৰ কাহিনী আৰু অভিনয় অতি সুন্দৰ হৈছে।"
result = classifier(sample_text)

print(result)

---

Training Hyperparameters

The training pipeline utilized an identical optimization schedule as the other pre-trained transformer blocks in this project:

Parameter	Value
Base Model	xlm-roberta-base
Total Epochs	4.0
Train Batch Size	16
Learning Rate	2e-05
Weight Decay	0.01
Warmup Ratio	0.1
Checkpoint Selection	Best Macro F1

---

Limitations & Edge Cases

• The "Curse of Multilinguality": Because XLM-R shares its parameter capacity across a massive global language footprint (100+ languages), its specific representation space for lower-resource scripts like Assamese is somewhat diluted.
• Morphological Subtexts & Discourse Markers: While the model leverages broad semantic structures very well, it struggles more than highly localized models (like MuRIL) with regional morphological subtexts, context-dependent negation patterns, and local contrastive discourse markers.

---

Citation & Contact

If you use this model or refer to this work in your research, please link back to this Hugging Face repository or the GitHub project repository. If you use this model, please cite the associated preprint:

Dev, C., Baishya, B., Dey, N., Dutta, A., & Dutta, S. (2026). Sentiment Analysis of Assamese Reviews using Transfer Learning Models. Preprint. https://doi.org/10.2139/ssrn.6357982