Intergenerational Test Generation for Natural Language Processing Applications

TL;DR

This paper introduces NLPLego, an automated, linguistically grounded test generation method for NLP applications that creates diverse, grammatically correct test cases to detect model errors effectively across multiple tasks.

Contribution

It presents a novel, general test generation framework based on sentence parsing that automates error detection in NLP models using seed sentences and grammatical assembly.

Findings

Successfully detected thousands of errors in state-of-the-art models

Achieved around 95.7% precision in error detection

Demonstrated effectiveness across multiple NLP tasks

Abstract

The development of modern NLP applications often relies on various benchmark datasets containing plenty of manually labeled tests to evaluate performance. While constructing datasets often costs many resources, the performance on the held-out data may not properly reflect their capability in real-world application scenarios and thus cause tremendous misunderstanding and monetary loss. To alleviate this problem, in this paper, we propose an automated test generation method for detecting erroneous behaviors of various NLP applications. Our method is designed based on the sentence parsing process of classic linguistics, and thus it is capable of assembling basic grammatical elements and adjuncts into a grammatically correct test with proper oracle information. We implement this method into NLPLego, which is designed to fully exploit the potential of seed sentences to automate the test generation. NLPLego disassembles the seed sentence into the template and adjuncts and then generates new sentences by assembling context-appropriate adjuncts with the template in a specific order. Unlike the taskspecific methods, the tests generated by NLPLego have derivation relations and different degrees of variation, which makes constructing appropriate metamorphic relations easier. Thus, NLPLego is general, meaning it can meet the testing requirements of various NLP applications. To validate NLPLego, we experiment with three common NLP tasks, identifying failures in four state-of-art models. Given seed tests from SQuAD 2.0, SST, and QQP, NLPLego successfully detects 1,732, 5301, and 261,879 incorrect behaviors with around 95.7% precision in three tasks, respectively.