Can ChatGPT Replace a Template-based Code Generator?

Abstract

This article examines whether a large language model (LLM) tool, such as ChatGPT, can replace a templatebased source code generator.To this end, we conducted an experiment in which we attempted to replace an existing templatebased DAO class generator (which creates entity classes and a repository for a specified database table) with a solution in which templates of target classes were presented to ChatGPT alongside the source model.We then instructed ChatGPT to generate new classes.A novelty in this work is an attempt at two-stage cooperation with ChatGPT: first we provide the pattern, then we fill it.The experiment proved that, at present, such a solution yields results that are neither predictable nor replicable, and successive attempts to execute the same commands returned wildly varying results.ChatGPT randomly recognises the rules that are present in templates, and complex instructions impact the generated results negatively.At present, classic code generation methods yield markedly superior results.

Ähnliche Arbeiten