1. Problem Statement
2. Methodology
   1. Datasets used
   2. Model
3. Running the code
   1. Training
   2. Inference
4. Authors

The goal is to predict answers to real questions about Wikipedia articles using chaii-1 dataset.

chaii-1 is a question answering dataset in Hindi and Tamil (without the use of translation). Given a context and question, the goal of question answering is to predict the answer to the question by selecting a span from the context. Consider the following example from the dataset:

This work is based on a Kaggle competition: chaii - Hindi and Tamil Question Answering

• official chaii-1 dataset [Link]
• MLQA/XQuAD for Hindi [Link]
• SQuAD translated to Tamil [Link]

All these datasets can be found in one place: [here]

We finetuned XLM-RoBERTa for QuestionAnswering from alon-albalak/xlm-roberta-large-xquad checkpoint.

For hyperparameters, look at args in train.py.

To train yourself, use the following command:

python train.py

While running the training script, the files mentioned in Section 2.i have to be placed under dataset/ directory (relative to the training script).

For inference you can use inference.ipynb.

You can either fine-tune your own version of the model (by following the steps mentioned in the previous section) or use our version which is available [here].

• Subhalingam D (subhalingamd)
• Rachit Jain (rachit-0032)

This README uses texts from the assignment problem document provided in the course.