IndT5: A Text-to-Text Transformer for 10 Indigenous Languages

  In this work, we introduce IndT5, the first Transformer language model for Indigenous languages. To train IndT5, we build IndCorpu, a new corpus for 10 Indigenous languages and Spanish.

 

IndT5

We train an Indigenous language model adopting the unified and flexible text-to-text transfer Transformer (T5) approach. T5 treats every text-based language task as a “text-to-text" problem, taking text format as input and producing new text format as output. T5 is essentially an encoder-decoder Transformer, with the encoder and decoder similar in configuration and size to a BERT_Base but with some architectural modifications. Modifications include applying a normalization layer before a sub-block and adding a pre-norm (i.e., initial input to the sub-block output).

IndCourpus

We build IndCorpus, a collection of 10 Indigeous languages and Spanish comprising 1.17GB of text, from both Wikipedia and the Bible.

Data size and number of sentences in monolingual dataset (collected from Wikipedia and Bible)

Target Language	Wiki Size (MB)	Wiki #Sentences	Bible Size (MB)	Bible #Sentences
Hñähñu	-	-	1.4	7.5K
Wixarika	-	-	1.3	7.5K
Nahuatl	5.8	61.1K	1.5	7.5K
Guarani	3.7	28.2K	1.3	7.5K
Bribri	-	-	1.5	7.5K
Rarámuri	-	-	1.9	7.5K
Quechua	5.9	97.3K	4.9	31.1K
Aymara	1.7	32.9K	5	30.7K
Shipibo-Konibo	-	-	1	7.9K
Asháninka	-	-	1.4	7.8K
Spanish	1.13K	5M	-	-
Total	1.15K	5.22M	19.8	125.3K

Github

More details about our model can be found here: https://github.com/UBC-NLP/IndT5

BibTex

@inproceedings{nagoudi-etal-2021-indt5,
    title = "{I}nd{T}5: A Text-to-Text Transformer for 10 Indigenous Languages",
    author = "Nagoudi, El Moatez Billah  and Chen, Wei-Rui  and Abdul-Mageed, Muhammad  and Cavusoglu, Hasan",
    booktitle = "Proceedings of the First Workshop on Natural Language Processing for Indigenous Languages of the Americas",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.americasnlp-1.30",
    doi = "10.18653/v1/2021.americasnlp-1.30",
    pages = "265--271"
}