Auto Code Comment Assessment for Online Judge using Word Embedding and Word Mover's Distance

Abstract

Comments in source code are a form of inline documentation created by programmers to help others understand the function of the program. The students of the basic programming subject need how to learn to write better code comments which can be difficulties for the lecturer assessing. Therefore, the author proposes an automatic source code comment assessment method for the online judge system with a corpus-based text similarity approach. Word2vec, GloVe, and fastText models will be used to train word vectors with the Indonesian Wikipedia Dump. The Similarities will be measured using Word Mover's Distance (WMD). Experiments were carried out using epoch variations during the training process. Spearman's rho correlation coefficient, mean average error (MAE), and performance measurements of each model will be compared. The methods with the proposed word embedding approach still provide not good results.