XLM RoBERTa for Czech+English Extractive Question Answering

This is the XLM-RoBERTa-large model with a head for extractive question answering trained on a combination of English SQuAD 1.1 and Czech SQAD 3.0 Question Answering datasets. For the Czech SQAD 3.0, original contexts (=whole Wikipedia websites) were limited to fit the RoBERTa's context window, excluding ~3% of the samples.

Intended uses & limitations

This model is purposed to extract a segment of a given context that contains an answer to a given question (Extractive Question Answering) in English and Czech. Given the fine-tuning on two languages and a good reported zero-shot cross-lingual applicability of other fine-tuned XLM-RoBERTas, the model will likely work on other languages as well, with a decay in quality.

Note that despite its size, English SQuAD has a variety of reported biases (see, e.g. L. Mikula (2022), Chap. 4.1).

Usage

Here is how to use this model to answer the question on a given context using 🤗 Transformers in PyTorch:

from transformers import AutoTokenizer, AutoModelForQuestionAnswering

tokenizer = AutoTokenizer.from_pretrained("gaussalgo/xlm-roberta-large_extractive-QA_en-cs")
model = AutoModelForQuestionAnswering.from_pretrained("gaussalgo/xlm-roberta-large_extractive-QA_en-cs")

context = """
Podle slovenského lidového podání byl Juro Jánošík obdařen magickými předměty (kouzelná valaška, čarovný opasek),
které mu dodávaly nadpřirozené schopnosti. Okrádal především šlechtice,
trestal panské dráby a ze svého lupu vyděloval část pro chudé, tedy bohatým bral a chudým dával.
"""
question = "Jaké schopnosti daly magické předměty Juro Jánošíkovi?"

inputs = tokenizer(question, context, return_tensors="pt")
outputs = model(**inputs)
start_position = outputs.start_logits[0].argmax()
end_position = outputs.end_logits[0].argmax()
answer_ids = inputs["input_ids"][0][start_position:end_position]

print("Answer:")
print(tokenizer.decode(answer_ids))

Training

The model has been trained using Adaptor library v0.1.5, in parallel on both Czech and English data, with the following parameters:

training_arguments = AdaptationArguments(output_dir="train_dir",
learning_rate=1e-5,
stopping_strategy=StoppingStrategy.ALL_OBJECTIVES_CONVERGED,
do_train=True,
do_eval=True,
warmup_steps=1000,
max_steps=100000,
gradient_accumulation_steps=30,
eval_steps=100,
logging_steps=10,
save_steps=1000,
num_train_epochs=30,
evaluation_strategy="steps")

You can find the full training script in train_roberta_extractive_qa.py, reproducible after a specific data preprocessing for Czech SQAD in parse_czech_squad.py