Genius1237/xlm-roberta-large-tydip

Multilingual Politeness Classification Model

This model is based on xlm-roberta-large and is finetuned on the English subset of the TyDiP dataset as discussed in the original paper here.

Languages

In the paper, this model was evaluated on English + 9 Languages (Hindi, Korean, Spanish, Tamil, French, Vietnamese, Russian, Afrikaans, Hungarian). Given the model's good performance and XLMR's cross lingual abilities, it is likely that this finetuned model can be used for more languages as well.

Evaluation

The politeness classification accuracy scores on 10 languages from the TyDiP test set are mentioned below.

lang	acc
en	0.892
hi	0.868
ko	0.784
es	0.84
ta	0.78
fr	0.82
vi	0.844
ru	0.668
af	0.856
hu	0.812

Usage

You can use this model directly with a text-classification pipeline

from transformers import pipeline

classifier = pipeline(task="text-classification", model="Genius1237/xlm-roberta-large-tydip")

sentences = ["Could you please get me a glass of water", "mere liye पानी का एक गिलास ले आओ "]

print(classifier(sentences))
# [{'label': 'polite', 'score': 0.9076159000396729}, {'label': 'impolite', 'score': 0.765066385269165}]

More advanced usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained('Genius1237/xlm-roberta-large-tydip')
model = AutoModelForSequenceClassification.from_pretrained('Genius1237/xlm-roberta-large-tydip')

text = "Could you please get me a glass of water"
encoded_input = tokenizer(text, return_tensors='pt')

output = model(**encoded_input)
prediction = torch.argmax(output.logits).item()

print(model.config.id2label[prediction])
# polite

Citation

@inproceedings{srinivasan-choi-2022-tydip,
    title = "{T}y{D}i{P}: A Dataset for Politeness Classification in Nine Typologically Diverse Languages",
    author = "Srinivasan, Anirudh  and
      Choi, Eunsol",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.findings-emnlp.420",
    doi = "10.18653/v1/2022.findings-emnlp.420",
    pages = "5723--5738",
    abstract = "We study politeness phenomena in nine typologically diverse languages. Politeness is an important facet of communication and is sometimes argued to be cultural-specific, yet existing computational linguistic study is limited to English. We create TyDiP, a dataset containing three-way politeness annotations for 500 examples in each language, totaling 4.5K examples. We evaluate how well multilingual models can identify politeness levels {--} they show a fairly robust zero-shot transfer ability, yet fall short of estimated human accuracy significantly. We further study mapping the English politeness strategy lexicon into nine languages via automatic translation and lexicon induction, analyzing whether each strategy{'}s impact stays consistent across languages. Lastly, we empirically study the complicated relationship between formality and politeness through transfer experiments. We hope our dataset will support various research questions and applications, from evaluating multilingual models to constructing polite multilingual agents.",
}