muril-large-cased

Overview

This model uses a BERT large architecture [1] pretrained from scratch using the Wikipedia [2], Common Crawl [3], PMINDIA [4] and Dakshina [5] corpora for 17 [6] Indian languages.

We use a training paradigm similar to multilingual bert, with a few modifications as listed:

• We include translation and transliteration segment pairs in training as well.
• We keep an exponent value of 0.3 and not 0.7 for upsampling, shown to enhance low-resource performance. [7]

See the Training section for more details.

Training

The MuRIL model is pre-trained on monolingual segments as well as parallel segments as detailed below :

• Monolingual Data : We make use of publicly available corpora from Wikipedia and Common Crawl for 17 Indian languages.
• Parallel Data : We have two types of parallel data :
  • Translated Data : We obtain translations of the above monolingual corpora using the Google NMT pipeline. We feed translated segment pairs as input. We also make use of the publicly available PMINDIA corpus.
  • Transliterated Data : We obtain transliterations of Wikipedia using the IndicTrans [8] library. We feed transliterated segment pairs as input. We also make use of the publicly available Dakshina dataset.

We keep an exponent value of 0.3 to calculate duplication multiplier values for upsampling of lower resourced languages and set dupe factors accordingly. Note, we limit transliterated pairs to Wikipedia only.

The model was trained using a self-supervised masked language modeling task. We do whole word masking with a maximum of 80 predictions. The model was trained for 1500K steps, with a batch size of 8192, and a max sequence length of 512.

Trainable parameters

All parameters in the module are trainable, and fine-tuning all parameters is the recommended practice.

Uses & Limitations

This model is intended to be used for a variety of downstream NLP tasks for Indian languages. This model is trained on transliterated data as well, a phenomenon commonly observed in the Indian context. This model is not expected to perform well on languages other than the ones used in pre-training, i.e. 17 Indian languages.

Evaluation

We provide the results of fine-tuning this model on a set of downstream tasks. We choose these tasks from the XTREME benchmark, with evaluation done on Indian language test-sets. All results are computed in a zero-shot setting, with English being the high resource training set language. The results for XLM-R (Large) are taken from the XTREME paper [9].

• Shown below are results on datasets from the XTREME benchmark (in %)

  PANX (F1)	bn	en	hi	ml	mr	ta	te	ur	Average
  XLM-R (large)	78.8	84.7	73.0	67.8	68.1	59.5	55.8	56.4	68.0
  MuRIL (large)	85.8	85.0	78.3	75.6	77.3	71.1	65.6	83.0	77.7

  UDPOS (F1)	en	hi	mr	ta	te	ur	Average
  XLM-R (large)	96.1	76.4	80.8	65.2	86.6	70.3	79.2
  MuRIL (large)	95.7	71.3	85.7	62.6	85.8	62.8	77.3

  XNLI (Accuracy)	en	hi	ur	Average
  XLM-R (large)	88.7	75.6	71.7	78.7
  MuRIL (large)	88.4	75.8	71.7	78.6

  XQUAD (F1/EM)	en	hi	Average
  XLM-R (large)	86.5/75.7	76.7/59.7	81.6/67.7
  MuRIL (large)	88.2/77.8	78.4/62.4	83.3/70.1

  MLQA (F1/EM)	en	hi	Average
  XLM-R (large)	83.5/70.6	70.6/53.1	77.1/61.9
  MuRIL (large)	84.4/71.7	72.2/54.1	78.3/62.9

  TyDiQA (F1/EM)	en	bn	te	Average
  XLM-R (large)	71.5/56.8	64.0/47.8	70.1/43.6	68.5/49.4
  MuRIL (large)	75.9/66.8	67.1/53.1	71.5/49.8	71.5/56.6

  The fine-tuning hyperparameters are as follows:

  Task	Batch Size	Learning Rate	Epochs	Warm-up Ratio
  PANX	32	2e-5	10	0.1
  UDPOS	64	5e-6	10	0.1
  XNLI	128	2e-5	5	0.1
  XQuAD	32	3e-5	2	0.1
  MLQA	32	3e-5	2	0.1
  TyDiQA	32	3e-5	3	0.1

References

[1]: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805, 2018.

[2]: Wikipedia

[3]: Common Crawl

[4]: PMINDIA

[5]: Dakshina

[6]: Assamese (as), Bengali (bn), English (en), Gujarati (gu), Hindi (hi), Kannada (kn), Kashmiri (ks), Malayalam (ml), Marathi (mr), Nepali (ne), Oriya (or), Punjabi (pa), Sanskrit (sa), Sindhi (sd), Tamil (ta), Telugu (te) and Urdu (ur).

[7]: Conneau, Alexis, et al. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116 (2019).

[8]: IndicTrans

[9]: Hu, J., Ruder, S., Siddhant, A., Neubig, G., Firat, O., & Johnson, M. (2020). [Xtreme: A massively multilingual multi-task benchmark for evaluating cross-l