niharikanath/assamese-sentiment-analysis

🌟 Assamese Sentiment Analysis with LSTM

Tags: #text-classification #sentiment-analysis #Assamese #LSTM

A deep learning-powered tool to classify Assamese text as Positive, Negative, or Neutral using an LSTM model tailored for the Assamese language.

🚀 Key Features

• —🔍 Sentiment Analysis for Assamese – Supports full sentiment classification of Assamese text
• —🧠 Deep Learning Backbone – Powered by TensorFlow/Keras with a Long Short-Term Memory (LSTM) network
• —✨ Advanced Preprocessing – Includes tokenization, text cleaning, optional stemming, and stopword removal
• —🧰 Custom Tokenization – Leverages AssameseTokenizer for accurate language handling
• —📈 Robust Evaluation Metrics – F1-score, precision, recall, and accuracy

🧠 Model Overview

🧪 Installation & Requirements

Clone the repo and install the requirements:

pip install -r requirements.txt

Install the custom Assamese tokenizer:

git clone https://github.com/KashyapKishore/AssameseTokenizer.git
cd AssameseTokenizer
pip install .

⚙️ Model Description

This model was developed using Assamese text data and trained with a custom tokenizer specifically designed for Assamese script. It uses an LSTM architecture, making it well-suited for capturing the sequence and context of natural language in sentiment classification tasks.

• —📚 Training Data The dataset was curated from public sources such as news articles, social media comments, and feedback forms, and was manually labeled into three sentiment classes: Positive, Neutral, and Negative.

• —🏋️ Training Procedure
• —✂️ Preprocessing: Text cleaning, tokenization using AssameseTokenizer, optional stemming and stopword removal

• —🔢 Input Handling: Sequences padded or truncated to a fixed length of 512 tokens

• —🧠 Architecture: Embedding layer → LSTM → Dense (Softmax)

• —💧 Regularization: Dropout layers to prevent overfitting

• —⚙️ Optimizer: Adam

• —🔁 Epochs: Trained for X epochs (replace with your actual number)

• —📊 Evaluation: Final validation accuracy and F1-score: Insert actual metrics here

📦 Intended Usage

Ideal for:

• —🗨️ Social media sentiment tracking in Assamese

• —📢 Public opinion & brand monitoring

• —📚 Research on low-resource NLP in Indic languages

• —⚠️ Limitations / Not Recommended For:

Code-mixed Assamese-English input

Domain-specific texts (e.g., legal, medical) without additional fine-tuning

🧪 Quickstart: Using the Model

You can load and run the model easily via Hugging Face's transformers pipeline:

from transformers import pipeline

model_name = "niharikanath/assamese-sentiment-analysis"
pipe = pipeline("text-classification", model=model_name, tokenizer=model_name)

result = pipe("এই খাবাৰটা একদম ভালো আছিল!")  # Sample Assamese sentence
print(result)

📚 Reference Citations

• —E. Grave*, P. Bojanowski*, P. Gupta, A. Joulin, T. Mikolov, Learning Word Vectors for 157 Languages

• —Assamese Tokenizer

🤝 In Collaboration with

• —Angshita Kashyap

• —Dhiraj Ballav Saikia

• —Pratyushee Talukdar