Reproduce “Neural Machine Translation for Mathematical Formulae” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Felix Petersen, Moritz Schubotz, Andre Greiner-Petter, Bela Gipp
Researchers from Stanford University, FIZ Karlsruhe, and the University of Göttingen developed neural machine translation systems for accurately translating between presentation (LaTeX) and content (Mathematica, Semantic LaTeX) representations of mathematical formulae. Their convolutional sequence-to-sequence model achieved 95.1% Exact Match accuracy for LaTeX to Mathematica translation, demonstrating a robust method for bridging this computational gap.
Key insights
- Convolutional sequence-to-sequence networks outperformed other neural architectures, including Transformers, for mathematical formula translation.
- Perplexity was not a reliable evaluation metric for mathematical language translation; Exact Match accuracy and Levenshtein Distance provided more meaningful insights.
- Preprocessing techniques like number substitution with placeholder tags were crucial for improving model performance in this domain.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.