An official website of the United States government
Although measuring held-out accuracy has been the primary approach to evaluate generalization, it often overestimates the performance of NLP models, while alternative approaches for evaluating models either focus on individual tasks or on specific behaviors. Inspired by principles of behavioral testing in software engineering, we introduce CheckList, a task-agnostic methodology for testing NLP models. CheckList includes a matrix of general linguistic capabilities and test types that facilitate comprehensive test ideation, as well as a software tool to generate a large and diverse number of test cases quickly. We illustrate the utility of CheckList with tests for three tasks, identifying critical failures in both commercial and state-of-art models. In a user study, a team responsible for a commercial sentiment analysis model found new and actionable bugs in an extensively tested model. In another user study, NLP practitioners with CheckList created twice as many tests, and found almost three times as many bugs as users without it.
more »
« less
Automated Testing Linguistic Capabilities of NLP Models
Natural language processing (NLP) has gained widespread adoption in the development of real-world applications. However, the black-box nature of neural networks in NLP applications poses a challenge when evaluating their performance, let alone ensuring it. Recent research has proposed testing techniques to enhance the trustworthiness of NLP-based applications. However, most existing works use a single, aggregated metric (i.e., accuracy) which is difficult for users to assess NLP model performance on fine-grained aspects, such as LCs. To address this limitation, we present ALiCT, an automated testing technique for validating NLP applications based on their LCs. ALiCT takes user-specified LCs as inputs and produces diverse test suite with test oracles for each of given LC. We evaluate ALiCT on two widely adopted NLP tasks, sentiment analysis and hate speech detection, in terms of diversity, effectiveness, and consistency. Using Self-BLEU and syntactic diversity metrics, our findings reveal that ALiCT generates test cases that are 190% and 2213% more diverse in semantics and syntax, respectively, compared to those produced by state-of-the-art techniques. In addition, ALiCT is capable of producing a larger number of NLP model failures in 22 out of 25 LCs over the two NLP applications.
more »
« less
- PAR ID:
- 10592632
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- ACM Transactions on Software Engineering and Methodology
- Volume:
- 33
- Issue:
- 7
- ISSN:
- 1049-331X
- Page Range / eLocation ID:
- 1 to 33
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Social media platforms are playing increasingly critical roles in disaster response and rescue operations. During emergencies, users can post rescue requests along with their addresses on social media, while volunteers can search for those messages and send help. However, efficiently leveraging social media in rescue operations remains challenging because of the lack of tools to identify rescue request messages on social media automatically and rapidly. Analyzing social media data, such as Twitter data, relies heavily on Natural Language Processing (NLP) algorithms to extract information from texts. The introduction of bidirectional transformers models, such as the Bidirectional Encoder Representations from Transformers (BERT) model, has significantly outperformed previous NLP models in numerous text analysis tasks, providing new opportunities to precisely understand and classify social media data for diverse applications. This study developed and compared ten VictimFinder models for identifying rescue request tweets, three based on milestone NLP algorithms and seven BERT-based. A total of 3191 manually labeled disaster-related tweets posted during 2017 Hurricane Harvey were used as the training and testing datasets. We evaluated the performance of each model by classification accuracy, computation cost, and model stability. Experiment results show that all BERT-based models have significantly increased the accuracy of categorizing rescue-related tweets. The best model for identifying rescue request tweets is a customized BERT-based model with a Convolutional Neural Network (CNN) classifier. Its F1-score is 0.919, which outperforms the baseline model by 10.6%. The developed models can promote social media use for rescue operations in future disaster events.more » « less
-
The low cost of resource ownership and flexibility have led users to increasingly port their applications to the clouds. To fully realize the cost benefits of cloud services, users usually need to reliably know the execution performance of their applications. However, due to the random performance fluctuations experienced by cloud applications, the black box nature of public clouds and the cloud usage costs, testing on clouds to acquire accurate performance results is extremely difficult. In this paper, we present a novel cloud performance testing methodology called PT4Cloud. By employing non-parametric statistical approaches of likelihood theory and the bootstrap method, PT4Cloud provides reliable stop conditions to obtain highly accurate performance distributions with confidence bands. These statistical approaches also allow users to specify intuitive accuracy goals and easily trade between accuracy and testing cost. We evaluated PT4Cloud with 33 benchmark configurations on Amazon Web Service and Chameleon clouds. When compared with performance data obtained from extensive performance tests, PT4Cloud provides testing results with 95.4% accuracy on average while reducing the number of test runs by 62%. We also propose two test execution reduction techniques for PT4Cloud, which can reduce the number of test runs by 90.1% while retaining an average accuracy of 91%. We compared our technique to three other techniques and found that our results are much more accurate.more » « less
-
Testing deep neural networks (DNNs) has garnered great interest in the recent years due to their use in many applications. Black-box test adequacy measures are useful for guiding the testing process in covering the input domain. However, the absence of input specifications makes it challenging to apply black-box test adequacy measures in DNN testing. The Input Distribution Coverage (IDC) framework addresses this challenge by using a variational autoencoder to learn a low dimensional latent representation of the input distribution, and then using that latent space as a coverage domain for testing. IDC applies combinatorial interaction testing on a partitioning of the latent space to measure test adequacy. Empirical evaluation demonstrates that IDC is cost-effective, capable of detecting feature diversity in test inputs, and more sensitive than prior work to test inputs generated using different DNN test generation methods. The findings demonstrate that IDC overcomes several limitations of white-box DNN coverage approaches by discounting coverage from unrealistic inputs and enabling the calculation of test adequacy metrics that capture the feature diversity present in the input space of DNNs.more » « less
-
A growing body of research has demonstrated the inability of NLP models to generalize compositionally and has tried to alleviate it through specialized architectures, training schemes, and data augmentation, among other approaches. In this work, we study a different approach: training on instances with diverse structures. We propose a model-agnostic algorithm for subsampling such sets of instances from a labeled instance pool with structured outputs. Evaluating on both compositional template splits and traditional IID splits of 5 semantic parsing datasets of varying complexity, we show that structurally diverse training using our algorithm leads to comparable or better generalization than prior algorithms in 9 out of 10 dataset-split type pairs. In general, we find structural diversity to consistently improve sample efficiency compared to random train sets. Moreover, we show that structurally diverse sampling yields comprehensive test sets that are a lot more challenging than IID test sets. Finally, we provide two explanations for improved generalization from diverse train sets: 1) improved coverage of output substructures, and 2) a reduction in spurious correlations between these substructures.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3672455
