AI4Bharat-IndicNLP Dataset

The AI4Bharat-IndicNLP dataset is an ongoing effort to create a collection of large-scale, general-domain corpora for Indian languages. Currently, it contains 2.7 billion words for 10 Indian languages from two language families. We share pre-trained word embeddings trained on these corpora. We create news article category classification datasets for 9 languages to evaluate the embeddings. We evaluate the IndicNLP embeddings on multiple evaluation tasks.

You can read details regarding the corpus and other resources HERE. We showcased the AI4Bharat-IndicNLP dataset at REPL4NLP 2020 (collocated with ACL 2020) (non-archival submission as extended abstract). You can see the talk here: VIDEO.

You can use the IndicNLP corpus and embeddings for multiple Indian language tasks. A comprehensive list of Indian language NLP resources can be found in the IndicNLP Catalog. For processing the Indian language text, you can use the Indic NLP Library.

Table of contents

• Text Corpora
• Word Embeddings
• IndicNLP News Article Classification Dataset
• Publicly available Classification Datasets
• Morphanalyzers
• Citing
• License
• Contributors
• Contact

Text Corpora

The text corpus for 12 languages.

Language	# News Articles*	Sentences	Tokens	Link
as	0.60M	1.39M	32.6M	link
bn	3.83M	39.9M	836M	link
en	3.49M	54.3M	1.22B	link
gu	2.63M	41.1M	719M	link
hi	4.95M	63.1M	1.86B	link
kn	3.76M	53.3M	713M	link
ml	4.75M	50.2M	721M	link
mr	2.31M	34.0M	551M	link
or	0.69M	6.94M	107M	link
pa	2.64M	29.2M	773M	link
ta	4.41M	31.5M	582M	link
te	3.98M	47.9M	674M	link

Note

• The vocabulary frequency files contain the frequency of all unique tokens in the corpus. Each line contains one word along with frequency delimited by tab.
• For convenience, the corpus is already tokenized using the IndicNLP tokenizer. You can use the IndicNLP detokenizer in case you want a detokenized version.

Pre-requisites

To replicate the results reported in the paper, training and evaluation scripts are provided.

To run these scripts, the following tools/packages are required:

• FastText
• MUSE

For Python packages to install, see requirements.txt

Word Embeddings

DOWNLOAD

Version 1

language	pa	hi	bn	or	gu	mr	kn	te	ml	ta
vectors	link	link	link	link	link	link	link	link	link	link
model	link	link	link	link	link	link	link	link	link	link

Training word embeddings

$FASTTEXT_HOME/build/fasttext skipgram \
	-epoch 10 -thread 30 -ws 5 -neg 10    -minCount 5 -dim 300 \
	-input $mono_path \
	-output $output_emb_prefix 

Evaluation on word similarity task

Evaluate on the IIIT-H Word Similarity Database: DOWNLOAD

The above mentioned link is a cleaned version of the same database found HERE.

Evaluation Command

python scripts/word_similarity/wordsim.py \
	<embedding_file_path> \
	<word_sim_db_path> \
	<max_vocab>

Evaluation on word analogy task

Evaluate on the Facebook word analogy dataset.

Evaluation Command

First, add MUSE root directory to Python Path

export PYTHONPATH=$PYTHONPATH:$MUSE_PATH

python  scripts/word_analogy/word_analogy.py \
    --analogy_fname <analogy_fname> \
    --embeddings_path <embedding_file_path> \
    --lang 'hi' \
    --emb_dim 300 \
    --cuda

IndicNLP News Article Classification Dataset

We used the IndicNLP text corpora to create classification datasets comprising news articles and their categories for 9 languages. The dataset is balanced across classes. The following table contains the statistics of our dataset:

Language	Classes	Articles per Class
Bengali	entertainment, sports	7K
Gujarati	business, entertainment, sports	680
Kannada	entertainment, lifestyle, sports	10K
Malayalam	business, entertainment, sports, technology	1.5K
Marathi	entertainment, lifestyle, sports	1.5K
Oriya	business, crime, entertainment, sports	7.5K
Punjabi	business, entertainment, sports, politics	780
Tamil	entertainment, politics, sport	3.9K
Telugu	entertainment, business, sports	8K

DOWNLOAD

Evaluation Command

python3 scripts/txtcls.py --emb_path <path> --data_dir <path> --lang <lang code>

Publicly available Classification Datasets

We also evaluated the IndicNLP embeddings on many publicly available classification datasets.

• ACTSA Corpus: Sentiment analysis corpus for Telugu sentences.
• BBC News Articles: Text classification corpus for Hindi documents extracted from BBC news website.
• IIT Patna Product Reviews: Sentiment analysis corpus for product reviews posted in Hindi.
• INLTK Headlines Corpus: Obtained from inltk project. The corpus is a collection of headlines tagged with their news category. Available for langauges: gu, ml, mr and ta.
• IIT Patna Movie Reviews: Sentiment analysis corpus for movie reviews posted in Hindi.
• Bengali News Articles: Contains Bengali news articles tagged with their news category.

We have created standard test, validation and test splits for the above mentioned datasets. You can download them to evaluate your embeddings.

DOWNLOAD

Evaluation Command

To evaluate your embeddings on the above mentioned datasets, first download them and then run the following command:

python3 scripts/txtcls.py --emb_path <path> --data_dir <path> --lang <lang code>

License

These datasets are available under original license for each public dataset.

Morphanalyzers

IndicNLP Morphanalyzers are unsupervised morphanalyzers trained with morfessor.

DOWNLOAD

Version 1

pa	hi	bn	or	gu	mr	kn	te	ml	ta

Training Command

## extract vocabulary from embedings file
zcat $embedding_vectors_path |  \
    tail -n +2 | \
    cut -f 1 -d ' '  > $vocab_file_path

## train morfessor 
morfessor-train -d ones \
        -S $model_file_path \
        --logfile  $log_file_path \
        --traindata-list $vocab_file_path \
        --max-epoch 10 

Citing

If you are using any of the resources, please cite the following article:

@article{kunchukuttan2020indicnlpcorpus,
    title={AI4Bharat-IndicNLP Corpus: Monolingual Corpora and Word Embeddings for Indic Languages},
    author={Anoop Kunchukuttan and Divyanshu Kakwani and Satish Golla and Gokul N.C. and Avik Bhattacharyya and Mitesh M. Khapra and Pratyush Kumar},
    year={2020},
    journal={arXiv preprint arXiv:2005.00085},
}

We would like to hear from you if:

• You are using our resources. Please let us know how you are putting these resources to use.
• You have any feedback on these resources.

License

IndicNLP Corpus is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Contributors

• Anoop Kunchukuttan
• Divyanshu Kakwani
• Satish Golla
• Gokul NC
• Avik Bhattacharyya
• Mitesh Khapra
• Pratyush Kumar

This work is the outcome of a volunteer effort as part of AI4Bharat initiative.

Contact

• Anoop Kunchukuttan (anoop.kunchukuttan@gmail.com)
• Mitesh Khapra (miteshk@cse.iitm.ac.in)
• Pratyush Kumar (pratyushk@cse.iitm.ac.in)