ICSE 2024
Fri 12 - Sun 21 April 2024 Lisbon, Portugal

This paper discusses the limitations of evaluating Masked Language Models (MLMs), particularly Encoder-Based Transformers, in code completion tasks. We highlight that solely relying on accuracy-based measurements may lead to an overestimation of models’ capabilities by neglecting the syntax rules of programming languages. To address these issues, we introduce a technique called SyntaxEval in which Syntactic Capabilities are used to enhance the evaluation of MLMs. SyntaxEval automates the process of masking elements in the model input based on their Abstract Syntax Trees (ASTs) node type and calculates the distance between ASTs after making predictions. We conducted a case study on two popular MLMs using data from GitHub repositories. Our results showed negative causal effects between the node types and MLMs’ accuracy. We conclude that MLMs under study fail to predict some syntactic capabilities from source code. However, a large-scale empirical study is needed to corroborate which capabilities are statistically learned by any MLM after controlling for confounding bias.

Thu 18 Apr

Displayed time zone: Lisbon change

14:00 - 15:30
LLM, NN and other AI technologies 4Research Track / Industry Challenge Track / New Ideas and Emerging Results at Pequeno Auditório
Chair(s): David Nader Palacio William & Mary
14:00
15m
Talk
Programming Assistant for Exception Handling with CodeBERT
Research Track
Yuchen Cai University of Texas at Dallas, Aashish Yadavally University of Texas at Dallas, Abhishek Mishra University of Texas at Dallas, Genesis Montejo University of Texas at Dallas, Tien N. Nguyen University of Texas at Dallas
14:15
15m
Talk
An Empirical Study on Noisy Label Learning for Program Understanding
Research Track
Wenhan Wang Nanyang Technological University, Yanzhou Li Nanyang Technological University, Anran Li Nanyang Technological University, Jian Zhang Nanyang Technological University, Wei Ma Nanyang Technological University, Singapore, Yang Liu Nanyang Technological University
Pre-print
14:30
15m
Talk
An Empirical Study on Low GPU Utilization of Deep Learning Jobs
Research Track
Yanjie Gao Microsoft Research, yichen he , Xinze Li Microsoft Research, Bo Zhao Microsoft Research, Haoxiang Lin Microsoft Research, Yoyo Liang Microsoft, Jing Zhong Microsoft, Hongyu Zhang Chongqing University, Jingzhou Wang Microsoft Research, Yonghua Zeng Microsoft, Keli Gui Microsoft, Jie Tong Microsoft, Mao Yang Microsoft Research
DOI Pre-print
14:45
15m
Talk
Using an LLM to Help With Code Understanding
Research Track
Daye Nam Carnegie Mellon University, Andrew Macvean Google, Inc., Vincent J. Hellendoorn Carnegie Mellon University, Bogdan Vasilescu Carnegie Mellon University, Brad A. Myers Carnegie Mellon University
15:00
15m
Talk
MissConf: LLM-Enhanced Reproduction of Configuration-Triggered Bugs
Industry Challenge Track
Ying Fu National University of Defense Technology, Teng Wang National University of Defense Technology, Shanshan Li National University of Defense Technology, Jinyan Ding National University of Defense Technolog, Shulin Zhou National University of Defense Technology, Zhouyang Jia National University of Defense Technology, Wang Li National University of Defense Technology, Yu Jiang Tsinghua University, Liao Xiangke National University of Defense Technology
File Attached
15:15
7m
Talk
XAIport: A Service Framework for the Early Adoption of XAI in AI Model Development
New Ideas and Emerging Results
Zerui Wang Concordia University, Yan Liu Concordia University, Abishek Arumugam Thiruselvi Concordia University, Wahab Hamou-Lhadj Concordia University, Montreal, Canada
DOI Pre-print
15:22
7m
Talk
Which Syntactic Capabilities Are Statistically Learned by Masked Language Models for Code?
New Ideas and Emerging Results
Alejandro Velasco William & Mary, David Nader Palacio William & Mary, Daniel Rodriguez-Cardenas , Denys Poshyvanyk William & Mary
Pre-print