INVOCMAP: MAPPING METHOD NAMES TO METHOD INVOCATIONS VIA MACHINE LEARNING

Hung Phan, Ali Jannesari

TL;DR: This paper proposes a theory of classifying Method Invocations by different abstraction levels and conducting a statistical approach for code completion from method name to method invocation.

Abstract: Implementing correct method invocation is an important task for software developers. However, this is challenging work, since the structure of method invocation can be complicated. In this paper, we propose InvocMap, a code completion tool allows developers to obtain an implementation of multiple method invocations from a list of method names inside code context. InvocMap is able to predict the nested method invocations which their names didn’t appear in the list of input method names given by developers. To achieve this, we analyze the Method Invocations by four levels of abstraction. We build a Machine Translation engine to learn the mapping from the first level to the third level of abstraction of multiple method invocations, which only requires developers to manually add local variables from generated expression to get the final code. We evaluate our proposed approach on six popular libraries: JDK, Android, GWT, Joda-Time, Hibernate, and Xstream. With the training corpus of 2.86 million method invocations extracted from 1000 Java Github projects and the testing corpus extracted from 120 online forums code snippets, InvocMap achieves the accuracy rate up to 84 in F1- score depending on how much information of context provided along with method names, that shows its potential for auto code completion.

Code: https://tinyurl.com/yxbrl3x7

Keywords: Statistical Machine Translation, Method Invocation, Auto Code Completion, Software Engineering