Semantic Parsing of Geometry Statements Using Supervised Machine Learning on Synthetic Data

EasyChair Preprint 6414

Abstract

In this extended abstract, we report on our ongoing work on the automated translation of high-school geometry statements into a formal language of syntax trees in first-order logic. We see this as the first step before translating both statements and proofs, and before widening the scope to parsing natural language mathematics in general. Our approach is based on Arsenal, a framework developed at SRI International for building domain-specific semantic parsers translating natural language to structured representations, namely expression trees. Arsenal trains a model (in this case, a sequence-to-sequence model) from synthetic datasets that it generates from the grammar of expression trees that it targets, which is particularly useful for domains where ground truth data is sparse or even nonexistent.

Keyphrases: geometry, semantic parsing, sequence-to-sequence model, synthetic data, type checking

@booklet{EasyChair:6414,
  author    = {Salwa Tabet Gonzalez and Stéphane Graham-Lengrand and Julien Narboux and Natarajan Shankar},
  title     = {Semantic Parsing of Geometry Statements Using Supervised Machine Learning on Synthetic Data},
  howpublished = {EasyChair Preprint 6414},
  year      = {EasyChair, 2021}}