An Empirical Study on the Bugs Found while Reusing Pre-trained Natural Language Processing Models

TL;DR

This study analyzes 984 bugs from 11 popular NLP pre-trained models to understand their types, causes, and impacts, revealing challenges like robustness issues, bias propagation, and resource consumption.

Contribution

It provides a comprehensive taxonomy of bugs in NLP model reuse, highlighting key issues and patterns to guide future bug reduction efforts.

Findings

Limited access to model internals affects robustness.

Input validation issues propagate bias.

High resource use leads to crashes.

Abstract

In NLP, reusing pre-trained models instead of training from scratch has gained popularity; however, NLP models are mostly black boxes, very large, and often require significant resources. To ease, models trained with large corpora are made available, and developers reuse them for different problems. In contrast, developers mostly build their models from scratch for traditional DL-related problems. By doing so, they have control over the choice of algorithms, data processing, model structure, tuning hyperparameters, etc. Whereas in NLP, due to the reuse of the pre-trained models, NLP developers are limited to little to no control over such design decisions. They either apply tuning or transfer learning on pre-trained models to meet their requirements. Also, NLP models and their corresponding datasets are significantly larger than the traditional DL models and require heavy computation. Such reasons often lead to bugs in the system while reusing the pre-trained models. While bugs in traditional DL software have been intensively studied, the nature of extensive reuse and black-box structure motivates us to understand the different types of bugs that occur while reusing NLP models? What are the root causes of those bugs? How do these bugs affect the system? To answer these questions, We studied the bugs reported while reusing the 11 popular NLP models. We mined 9,214 issues from GitHub repositories and identified 984 bugs. We created a taxonomy with bug types, root causes, and impacts. Our observations led to several findings, including limited access to model internals resulting in a lack of robustness, lack of input validation leading to the propagation of algorithmic and data bias, and high-resource consumption causing more crashes, etc. Our observations suggest several bug patterns, which would greatly facilitate further efforts in reducing bugs in pre-trained models and code reuse.