/nlp-for-manglish

State of the Art Language models and Classifier for Code mixed Manglish (Malayalam and English) - spoken in Indian sub-continent.

Primary LanguageJupyter NotebookMIT LicenseMIT

NLP for Manglish (Code mixed Malayalam+English)

This repository contains state of the art Language models and Classifier for Code mixed Manglish (Malayalam and English) - spoken in Indian sub-continent.

Dataset

  1. Malayalam Wikipedia Articles : Preprocessed, Transliterated and Translated versions of this dataset, used for language modeling in this repo, can be downloaded directly from here

  2. Dravidian Codemix HASOC @ FIRE 2020

  3. Dravidian Codemix Sentiment Analysis @ FIRE 2020

Results

Language Model Perplexity (on validation set)

Architecture/Dataset Malayalam Wikipedia Articles Malayalam Wikipedia Articles
Latin Script Mixed Script
ULMFiT 45.84 41.22

Classification Metrics

ULMFiT
Dataset F1 Precision Recall Notebook to Reproduce results
Dravidian Codemix HASOC @ FIRE 2020 (Latin Script) 0.74 0.76 0.72 Link
Dravidian Codemix HASOC @ FIRE 2020 (Mixed Script) 0.91 0.92 0.91 Link
Dravidian Codemix Sentiment Analysis @ FIRE 2020 0.69 0.69 0.7 Link

Visualizations

Word Embeddings
Architecture/Dataset Malayalam Wikipedia Articles Malayalam Wikipedia Articles
Latin Script Mixed Script
ULMFiT Embeddings projection Embeddings projection

Pretrained Models

Language Models

Download Latin script pretrained ULMFiT LM from here

Download Mixed Script pretrained ULMFiT LM from here

Tokenizer

Trained tokenizer using Google's sentencepiece

Download the trained model and vocabulary for both Latin and Mixed scripts from here