chaii - Challenge in AI for India 2021

With nearly 1.4 billion people, India is the second-most populated country in the world. Yet Indian languages, like Hindi and Tamil, are underrepresented on the web. Popular Natural Language Understanding (NLU) models perform worse with Indian languages compared to English, the effects of which lead to sub-par experiences in downstream web applications for Indian users. With more attention from the Kaggle community and your novel machine learning solutions, we can help Indian users make the most of the web.

Predicting answers to questions is a common NLU task, but not for Hindi and Tamil. Current progress on multilingual modeling requires a concentrated effort to generate high-quality datasets and modelling improvements. Additionally, for languages that are typically underrepresented in public datasets, it can be difficult to build trustworthy evaluations. We hope the training data provided for this competition—and other datasets generated by participants—will enable future machine learning for Indic languages. 

About the Contest 

In this Kaggle competition, the goal for participants is to predict answers to real questions about Wikipedia articles. chaii-1, a new dataset provided by Google Research India, covers two Indian languages (Hindi and Tamil) with question-answer pairs. We additionally provide a baseline model.

Participants are highly encouraged to supplement the dataset by creating their own additional datasets as well. We will award the top five teams, USD 2000 each, that can achieve stellar performance on the test dataset. 

chaii-1 provides a realistic information-seeking task with questions written by expert data annotators who want to know the answer but do not have the answer. Data is collected directly from each language without the use of translation and was annotated by native speakers.