Towards Usable API Documentation

The learning and usage of an API is supported by documentation. Like source code, API documentation is itself a software product. Several research results show that bad design in API documentation can make the reuse of API features difficult. Indeed, similar to code smells, poorly designed API documentation can also exhibit 'smells'. Such documentation smells can be described as bad documentation styles that do not necessarily produce incorrect documentation but make the documentation difficult to understand and use. This thesis aims to enhance API documentation usability by addressing such documentation smells in three phases. In the first phase, we developed a catalog of five API documentation smells consulting literature on API documentation issues and online developer discussion. We validated their presence in the real world by creating a benchmark of 1K official Java API documentation units and conducting a survey of 21 developers. The developers confirmed that these smells hinder their productivity and called for automatic detection and fixing. In the second phase, we developed machine-learning models to detect the smells using the 1K benchmark, however, they performed poorly when evaluated on larger and more diverse documentation sources. We explored more advanced models; employed re-training and hyperparameter tuning to further improve the performance. Our best-performing model, RoBERTa, achieved F1-scores of 0.71-0.93 in detecting different smells. In the third phase, we first focused on evaluating the feasibility and impact of fixing various smells in the eyes of practitioners. Through a second survey of 30 practitioners, we found that fixing the lazy smell was perceived as the most feasible and impactful. However, there was no universal consensus on whether and how other smells can/should be fixed. Finally, we proposed a two-stage pipeline for fixing lazy documentation, involving additional textual description and documentation-specific code example generation. Our approach utilized a large language model, GPT- 3, to generate enhanced documentation based on non-lazy examples and to produce code examples. The generated code examples were refined iteratively until they were error-free. Our technique demonstrated a high success rate with a significant number of lazy documentation instances being fixed and error-free code examples being generated