In last few years, applying language model to source code is the state-of-the-art method for solving the problem of code completion. However, compared with natural language, code has more obvious repetition characteristics. For example, a variable can be used many times in the following code. Variables in source code have a high chance to be repetitive. Cloned code and templates, also have the property of token repetition. Capturing the token repetition of source code is important. In different projects, variables or types are usually named differently. This means that a model trained in a finite data set will encounter a lot of unseen variables or types in another data set. How to model the semantics of the unseen data and how to predict the unseen data based on the patterns of token repetition are two challenges in code completion. Hence, in this paper, token repetition is modelled as a graph, we propose a novel REP model which is based on deep neural graph network to learn the code toke repetition. The REP model is to identify the edge connections of a graph to recognize the token repetition. For predicting the token repetition of token [Formula: see text], the information of all the previous tokens needs to be considered. We use memory neural network (MNN) to model the semantics of each distinct token to make the framework of REP model more targeted. The experiments indicate that the REP model performs better than LSTM model. Compared with Attention-Pointer network, we also discover that the attention mechanism does not work in all situations. The proposed REP model could achieve similar or slightly better prediction accuracy compared to Attention-Pointer network and consume less training time. We also find other attention mechanism which could further improve the prediction accuracy.
Multi-task learning based pre-trained language model for code completion