NMT for Finno-Ugric Languages

This is an NMT system for translating between Võro, Livonian, North Sami, South Sami as well as Estonian, Finnish, Latvian and English. It was created by fine-tuning Facebook's m2m100-418M on the liv4ever and smugri datasets.

Tokenizer

Four language codes were added to the tokenizer: liv, vro, sma and sme. Currently the m2m100 tokenizer loads the list of languages from a hard-coded list, so it has to be updated after loading; see the code example below.

Usage example

Install the transformers and sentencepiece libraries: pip install sentencepiece transformers


from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("tartuNLP/m2m100_418M_smugri")
#Fix the language codes in the tokenizer
tokenizer.id_to_lang_token = dict(list(tokenizer.id_to_lang_token.items()) + list(tokenizer.added_tokens_decoder.items()))
tokenizer.lang_token_to_id = dict(list(tokenizer.lang_token_to_id.items()) + list(tokenizer.added_tokens_encoder.items()))
tokenizer.lang_code_to_token = { k.replace("_", ""): k for k in tokenizer.additional_special_tokens }
tokenizer.lang_code_to_id = { k.replace("_", ""): v for k, v in tokenizer.lang_token_to_id.items() }

model = AutoModelForSeq2SeqLM.from_pretrained("tartuNLP/m2m100_418M_smugri")

tokenizer.src_lang = 'liv'

encoded_src = tokenizer("Līvõ kēļ jelāb!", return_tensors="pt")

encoded_out = model.generate(**encoded_src, forced_bos_token_id = tokenizer.get_lang_id("sme"))
print(tokenizer.batch_decode(encoded_out, skip_special_tokens=True))

The output is Livčča giella eallá.