Classifying and recommending reference API documentation
Yam Chhetri , Former Master's student, SOCS McGill

Area: software evolution and natural language processing

Abstract:

Reference documentation is an important source of information on API usage. However, information useful to programmers can be buried in irrelevant or boiler-plate text, making it difficult to discover. In this talk, I will explain a technique we used to detect and recommend fragments of API documentation potentially important to a programmer who has already decided to use a certain API element. We categorize text fragments in API documentation based on whether they contain information that is indispensable, valuable, or neither. From the fragments that contain knowledge worthy of recommendation, we extract word patterns, and use these patterns to automatically find new fragments that contain similar knowledge in unseen documentation. In an evaluation study with randomly-sampled method definitions from ten open source systems, we found that with a training set of about 1 000 manually-classified sentences, we could issue recommendations with about, on average, 90% precision and 69% recall.