Machine Translation of Mathematical Text

We have implemented a machine translation system, the PolyMath Translator, for LaTeX documents containing mathematical text. The current implementation translates English to French LaTeX, attaining BLEU score of 53.6 on held-out test corpus sentences. It produces that can be compiled PDF without further editing. system first converts body an input document into sentences math tokens, using pandoc universal converter parse input. trained Transformer-based translator model, OpenNMT, combined small proportion domain-specific Our full uses this Transformer model and also Google Translate with custom glossary, latter being used as backup better handle linguistic features do not appear in our training dataset. is when does confidence its translation, determined by high perplexity score. Ablation testing demonstrates tokenization symbolic expressions essential quality translations produced system. published Translator available web service at http://www.polymathtrans.ai.