Metric-Type Identification for Multilevel Header Numerical Tables in Scientific Papers

Numerical tables are widely used to present experimental results in scientific papers. For table understanding, a metric-type is essential discriminate numbers the tables. We introduce new information extraction task, identification from multi-level header numerical tables, and provide dataset extracted papers consisting of captions, metric-types. then propose two joint-learning neural classification generation schemes featuring pointer-generator-based BERT-based models. Our show that joint models can handle both in-header out-of-header problems.