Can Machines Read Coding Manuals Yet? -- A Benchmark for Building Better Language Models for Code Understanding

TL;DR

This paper introduces BLANCA, a benchmark suite for evaluating language models on code-related textual artifacts, revealing current limitations and improvements through fine-tuning and multi-task training.

Contribution

The paper presents BLANCA, the first systematic benchmark for assessing language models on code documentation and forum discussions, and demonstrates how fine-tuning enhances their performance.

Findings

Fine-tuning significantly improves model performance on BLANCA tasks.

Multi-task training over BLANCA tasks leads to better code understanding models.

Current state-of-the-art models still have room for improvement on code-related text understanding.

Abstract

Code understanding is an increasingly important application of Artificial Intelligence. A fundamental aspect of understanding code is understanding text about code, e.g., documentation and forum discussions. Pre-trained language models (e.g., BERT) are a popular approach for various NLP tasks, and there are now a variety of benchmarks, such as GLUE, to help improve the development of such models for natural language understanding. However, little is known about how well such models work on textual artifacts about code, and we are unaware of any systematic set of downstream tasks for such an evaluation. In this paper, we derive a set of benchmarks (BLANCA - Benchmarks for LANguage models on Coding Artifacts) that assess code understanding based on tasks such as predicting the best answer to a question in a forum post, finding related forum posts, or predicting classes related in a hierarchy from class documentation. We evaluate the performance of current state-of-the-art language models on these tasks and show that there is a significant improvement on each task from fine tuning. We also show that multi-task training over BLANCA tasks helps build better language models for code understanding.