hindi2vec

State-of-the-Art Language Modeling and Text Classification in Hindi Language

Results

We achieved State of the Art Perplexity = 46.81 for Hindi compared to 40.68 for English (lower is better)

• To the best of my knowledge on September 18, 2018

Update: nlp-for-hindi uses sentencepiece instead of the word based spacCy tokenizer which I use. On those tokens, the measured perplexity for that LM is ~35. I encourage you to check that work out as well.

Downloads

• Pretrained Language Models that you can use in your classification for transfer learning
• EXCLUSIVE: BBC Hindi data of 4335 documents for text classification and text summarization. Release Notes
• Raw Data for Language Model shared above: Hindi Wikipedia with about 21k unique tokens for minfreq = 50
  • Wikipedia Processed Data - please use this to train your model

TODO

• Language modeling based on wikipedia dump
• Release Language Models: Hindi Language Model
• Create Text classification Datasets: BBC Hindi
• Benchmark text classification with FastText

Idea Dump

• Change the custom head to be used for transliteration instead of classification, Hindi script (Devnagri) to English script (Roman)
• MTL tasks for training and inference using custom heads
• Text to Speech - using datasets from news recordings or Hindi subtitles of dubbed movies

FastAI Installation

This version of the notebook uses fastai lib's v0.7, used in their Part 2 v2 course in Summer 2018. The best way to install it via conda as mentioned here

Special thanks to Jeremy, Rachel and other contributors to fastai. This work is a reproduction of their work in English to Hindi. Thanks to @cstorm125 for thai2vec which inspired this work.