Learning Transfers over Several Programming Languages

• DOI: 10.48550/arxiv.2310.16937
• Publication Date: 2023-01-01
• Authors: Razan Baltaji; Saurabh Pujar; Louis Mandel; Martin Hirzel; Luca Buratti; Lav R. Varshney
• Keywords: Third-generation programming language

Large language models (LLMs) have become remarkably good at improving developer productivity for high-resource programming languages. These use two kinds of data: large amounts unlabeled code samples pre-training and relatively smaller labeled fine-tuning or in-context learning. Unfortunately, many languages are low-resource, lacking most tasks often even samples. Therefore, users low-resource (e.g., legacy new languages) miss out on the benefits LLMs. Cross-lingual transfer uses data from a source to improve model performance target language. It has been well-studied natural languages, but received little attention This paper reports extensive experiments four using transformer-based LLM 11 41 explore following questions. First, how well does cross-lingual work given task across different pairs. Second, language, should one choose Third, which characteristics pair predictive performance, that depend task. Our empirical study with 1,808 reveals practical scientific insights, such as Kotlin JavaScript being transferable relying substantially features. Overall, we find learning transfers several