Transduce: learning transduction grammars for string transformation

TL;DR

Transduce tackles automatic synthesis of string transformation programs from input-output examples by introducing abstract transduction grammars that encode transformations as Prolog clauses. The method proceeds through learning (construction, abstraction, encoding as two integer sequences, and compressive generalization) and inference (rule construction for a given input length and execution on the input), enabling efficient learning of positional transformations with minimal examples. Empirically, it outperforms the state of the art on positional tasks and handles insertions and deletions, while highlighting limitations related to contextual knowledge and readability. The work advances program-synthesis research by reducing inductive bias through grammar-based generalization and points to future improvements in handling broader contexts and numeric data.

Abstract

The synthesis of string transformation programs from input-output examples utilizes various techniques, all based on an inductive bias that comprises a restricted set of basic operators to be combined. A new algorithm, Transduce, is proposed, which is founded on the construction of abstract transduction grammars and their generalization. We experimentally demonstrate that Transduce can learn positional transformations efficiently from one or two positive examples without inductive bias, achieving a success rate higher than the current state of the art.