PatiGonit22K: A Comprehensive Dataset for Solving Complex Bengali MWPs
This paper introduces PatiGonit22K, a comprehensive and culturally adapted dataset of 22,441 annotated Bengali Mathematical Word Problems designed to advance research in mathematical reasoning and educational NLP for low-resource languages.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where computers are like brilliant, hungry students who can read almost any language on Earth, but they struggle when it comes to solving math word problems in languages they haven't studied much. This is the playground of "Natural Language Understanding," a branch of science where we teach machines to not just recognize words, but to understand the stories hidden inside them and the numbers those stories describe. For a long time, these computer students have had plenty of practice books in English and Chinese, but for languages like Bengali, spoken by hundreds of millions of people, the library was nearly empty. Without enough practice problems, the computers couldn't learn how to reason through tricky questions like, "If Rachel buys 7 chairs and 3 tables, and it takes 4 minutes to assemble each, how long does the whole job take?" This paper steps into that empty library to build a massive new practice book, specifically designed to help computers get smarter at solving math puzzles in Bengali.
The authors of this paper, a team of researchers from universities in Dhaka, Bangladesh, noticed that while a previous collection of Bengali math problems existed, it was a bit too simple. It was like a gym with only light dumbbells; it helped beginners, but it couldn't train the computer muscles needed for heavy lifting. To fix this, they introduced PatiGonit22K, a new, expanded dataset containing 22,441 math word problems. Think of this dataset as a giant, carefully curated treasure chest of riddles. The team didn't just copy-paste; they took existing problems, translated them with extreme care to ensure the Bengali words felt natural and culturally correct, and then added a huge batch of new, complex challenges.
The real magic of PatiGonit22K is its mix of difficulty levels. The dataset is split into two main groups: 5,412 "simple" problems that involve just one math operation (like only adding or only subtracting), and a whopping 17,029 "complex" problems that require juggling multiple operations at once (like adding, then multiplying, then dividing). This balance is crucial because it allows researchers to test if a computer can handle a basic arithmetic task or if it can get lost when the problem gets messy. The researchers verified every single problem with a team of bilingual experts to make sure the math was right and the language was clear, removing any confusing or duplicate entries.
What does this mean for the future? The paper suggests that by providing this larger, more diverse collection of problems, researchers can finally train and test artificial intelligence models to see how well they really understand Bengali math. Previous studies had shown that models could solve simple Bengali problems with high accuracy, but this new dataset pushes the boundaries to see if they can tackle the harder, multi-step puzzles that are common in real life. The authors present this dataset as a robust new benchmark, a standard ruler against which future AI models can be measured. They don't claim to have solved the problem of AI reasoning forever, but they have handed the scientific community a much better set of training wheels and a tougher obstacle course to ensure that the next generation of AI can truly think through math problems in Bengali, not just guess the answer.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.