This is a seq2seq model (decoder half) to "flip" gender in first-person Arabic sentences.
The model can augment your existing Arabic data, or generate counterfactuals
to test a model's decisions (would changing the gender of the subject or speaker change output?).
Intended Examples:
'أنا سعيد' <=> 'انا سعيدة'
'ركض إلى المتجر' <=> 'ركضت إلى المتجر'
People's names, gender pronouns, gendered words (father, mother), and many other values are currently unchanged by this model. Future versions may be trained on more data.
Sample Code
import torch
from transformers import AutoTokenizer, EncoderDecoderModel
model = EncoderDecoderModel.from_encoder_decoder_pretrained(
"monsoon-nlp/ar-seq2seq-gender-encoder",
"monsoon-nlp/ar-seq2seq-gender-decoder",
min_length=40
)
tokenizer = AutoTokenizer.from_pretrained('monsoon-nlp/ar-seq2seq-gender-decoder') # same as MARBERT original
input_ids = torch.tensor(tokenizer.encode("أنا سعيدة")).unsqueeze(0)
generated = model.generate(input_ids, decoder_start_token_id=model.config.decoder.pad_token_id)
tokenizer.decode(generated.tolist()[0][1 : len(input_ids[0]) - 1])
> 'انا سعيد'
The Arabic model encoder and decoder started with weights and vocabulary from
MARBERT from UBC-NLP,
and was trained on the
Arabic Parallel Gender Corpus
from NYU Abu Dhabi. The text is first-person sentences from OpenSubtitles, with parallel
gender-reinflected sentences generated by Arabic speakers.
This model is useful to generate male and female text samples, but falls
short of capturing gender diversity in the world and in the Arabic
language. This subject is discussed in the bias statement of the
Gender Reinflection paper.
