Unsupervised Sentiment Analysis for Code-mixed Data

We use embeddings techniques like MUSE, LASER, XLM, MutltiBPEemd, fasttext to efficiently transfer knowledge from monolingual test to code-mix text for sentiment analysis of code-mixed text. More information about the methods tried here can be found in here.

Environment

All the dependencies of the code are listed in requirements.txt.

pip

    pip install -r requirements.txt
    PYTHONIOENCODING=utf-8 python -m laserembeddings download-models

docker

    # build the image 
    docker build -t unsacmt .
    
    # run the container
    nvidia-docker run -v $PWD:/app -p 8989:8989 unsacmt
    
    # launch a jupyter notebook
    jupyter notebook --ip 0.0.0.0 --port 8989 --allow-root

Data

The Sentiment Analysis data is present is data/cm/.
The custom fastText embedding is provided here. # TODO
The aligned MUSE embedding is provided here. # TODO

Files Description

• notebooks/archive/*.ipynb: old notebooks with many more experiments than mentioned in the paper.
• notebooks/Results.ipynb: a notebook with all the experiments
• src/utills.py: code for reading raw data and f1 score
• src/trainer.py: code for following training curriculum given the model and data
• src/models.py: code for simple neural network models used by use
• src/data_prep.py: code for applying different kinds of embeddings on sentiment analysis dataset

Citation

@misc{yadav2020unsupervised,
    title={Unsupervised Sentiment Analysis for Code-mixed Data},
    author={Siddharth Yadav and Tanmoy Chakraborty},
    year={2020},
    eprint={2001.11384},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}