Al-Driven Automated Software Documentation Generation for Enhanced Development Productivity

Saransh Arora

Abstract:

In the competitive software development
industry, effective and superior documentation is a must today.
Based on complicated AI models, automated code generation
solves this problem and helps produce documentation easily.
This work introduces a new approach to automatically
generating software documentation, focusing on fine-tuning
sophisticated AI models such as GPT-2 and RoBERTa by
leveraging a large existing dataset from the GitHub
CodeSearchNet challenge. The researchers indicate that
RoBERTa outperforms GPT-2 on both accuracy and loss
metrics, with an amazing accuracy score of 99.94% vs 74.37%
for GPT-2. RoBERTa also demonstrates much lower training
and validation losses to highlight its advantages. Another benefit
of RoBERTa is its significantly smaller training and validation
losses (0.010 and 0.002, respectively) than GPT-2 (1.407 and
1.268). The implication of the above is that quality of
documentation and more efficient development are achievable
with AI-driven automated documentation production.