Predicting Structure in Handwritten Algebra Data From Low Level Features

Background A goal in computer vision and natural language processing is to predict high level structure over handwritten algebra data. This is typically done by clustering or otherwise identifying symbols, then applying a known or learned grammar to generate dependency trees. However, these methods are often not generic and require domain knowledge. Aim The goal of this paper is to determine whether it is practical to predict dependency trees from low level features of handwritten algebra data without the use of symbol identification, and to determine which algorithms are best suited for this task. Data Our data consists of 400 pages of handwritten algebra data collected by one of our researchers. Methods We use transition-based parsing to generate dependency trees, and various no-regret online imitation learning algorithms to predict transitions. Results We find that a simple algorithm called greedy transition-based parsing, which has regretful learning and greedy decoding, gives similar or better results compared to the no-regret algorithms. This is surprising, as theory suggests noregret algorithms should outperform regretful algorithms. We discuss possible explanations for this discrepancy. Conclusions We conclude that it is practical to predict dependency trees over handwritten algebra data without the use of symbol identification.