Two-Dimensional Encoding Method for Neural Synthesis of Tabular Transformation by Example

Abstract

Programming by example (PBE) is a technology that makes data transformation tasks, especially tabular data transformation, easier for data analysts by automatically generating transformation programs from user-given input–output examples. In recent years, PBE research using machine learning (ML) has emerged because of the recent success of ML in various research fields. We developed an ML-based PBE system for tabular data transformation in previous work. The system is based on the Transformer model that fits sequential data and not two-dimensional structured data like tabular data. Inspired by recent work applying a Transformer model to tasks using two-dimensional data, such as the query answering task for tables or the image computer vision task, we propose a Transformer-based model with positional encoding for two-dimensional tabular data, called tabular positional encoding, to improve the Transformer-based model developed in our previous work. We implemented our proposed model and conducted various experiments. The experimental results show that the Transformer-based model with tabular positional encoding achieves much higher performance than our previous work.