API parameter recommendation based on language model and program analysis

Can an old fashioned feature extraction and a light-weight model improve vulnerability type identification performance?

“…Finally, 𝜇VulDeePecker is designed for the function call vulnerabilities, while our work has no limit to the kind of Learning-based approaches for SE tasks. Several studies have been proposed for specific SE tasks, including code suggestion/completion [46,47,48,49,50], program synthesis [51], pull request description generation [52,53], code summarization [54,55,56], code clones [57], fuzz testing [58], code-text translation [59], and program repair [60,61]. Recently, several learning techniques have been proposed to learn representing source code for specific SE applications [62,37] or general SE tasks [29,38,63].…”