Abstract:

In the regularly held Trends in International Mathematics and Science Study (TIMSS) and Programme for International Student Assessment (PISA) surveys, Singapore has performed consistently well. Singapore Math is a teaching approach originally developed by the country's Ministry of Education in the 1980s for its public schools. Since then, Singapore Math has been widely adopted in various forms around the world over the past twenty years.This paper describes an independent research project conducted by a pair of high school students under the mentorship of a senior research scientist from the National Institute of Education, Singapore, from April 2022 till March 2023. It describes a deep neural solver which the students designed and trained to solve Singapore mathematics word problems and to provide equations as explanation. In contrast to common approaches of plainly using the Math23K dataset, we translated the large dataset into English and inserted Singaporean Mathematics word problems that came from test papers and assessment books. By using an Encoder-Decoder model which uses recurrent neural networks (RNN), we fed the processed datasets into the model and evaluated the performance of the model based on different question types. Within the timeframe mandated by the research project, we were able to achieve an accuracy of the model for the Math23K dataset of 37.4%. The accuracy was lower for Singapore Mathematical word problems. The lower accuracy can be attributed to the model learning mainly Mathematical problems from a dataset derived from a non-Singaporean context, with the consequence that it was not able to sufficiently identify the new question types. The model performed best with “More than, Less than, As many as” questions, achieving an accuracy of 35%.

Keywords:

Encoder-decoder model, Singapore primary school, mathematical word problems, artificial intelligence, machine learning, neural networks, accuracy, loss.