NeSy is alive and well: A LLM-driven symbolic approach for better code comment data generation and classification

The era of Large Language Models (LLMs) has introduced agents capable of handling different tasks and performing well on them in domains such as text, image and audio [1]. A popular extension to the use of LLMs is in applying them to other data formats often used by humans in their daily activities. One such data source is code which circulates heavily and makes up a crucial block of technological projects [2]. The public availability of code-hosting repositories like GitHub on the web makes code an accessible data source and a valuable input for LLMs to tackle code-related challenges [2]. These tasks can range from identifying correct code to generating entirely new source code [2]. This has made source code datasets an invaluable part of the pre-training of modern LLM agents [2].