LSTM-Based with Deterministic Negative Sampling for API Suggestion

Modern programming relies on a large number of fundamental APIs, but programmers often take great effort to remember names and the usage of APIs when coding, and repeatedly search the related API documents or Q&A websites (e.g. Stack Overflow). To improve the programming efficiency, we present a Java API suggestion model called APIHelper which learns API sequence pattern via the Long Short-Term Memory (LSTM) network, then provides API suggestion based on the program context. Comparing with statistical methods (e.g. Hidden Markov Model (HMM), N-gram), which require establishing one specific model for each class, we propose Deterministic Negative Sampling (DNS) to make API suggestion for a large number of Java classes by one single end-to-end LSTM. To verify this approach, we make API suggestion for 50,000 Java classes and evaluate it with accuracy and top-K accuracy. The results show that APIHelper outperforms other research works both on accuracy and computation efficiency.