Abstract

In the USA, math proficiency levels these days are lower than ever before, which is problematic, since math is commonly used throughout life and math enables people to better solve problems, understand patterns, quantify relationships, and make predictions of the future. While Math Information Retrieval (MIR) as an area of study is relatively new, it is essential and provides a means to search for relevant sources of math information to those who are studying math, something which is difficult to do for people without prior knowledge of a specific math subject area they are looking for. In order to develop a robust MIR system, designers must be able to process Math Equations (MEs) to a format that the system can use, which is difficult due to various formats math information are stored in, including visual images and document texts. In solving this problem, we propose a ME extraction system that (i) applies a one shot object detector to identify math equations in digital images using an efficient neural architecture search method and (ii) employs a Sequence-to-Sequence (Seq2Seq) encoder-decoder system to recognize math equation symbols based on the Bayesian Neural Network (BNN) row encoding. The proposed system balances speed and accuracy of a Math Information Retrieval (IR) system.