Towards Effectively Testing Sequence-to-Sequence models from White-Box Perspectives
Bibliographic record
Abstract
In the field of Natural Language Processing (NLP), which encompasses diverse tasks such as machine translation, question answering, and others, there has been notable advancement in recent years. Despite this progress, NLP systems, including those based on sequence-to-sequence models, confront various challenges. To tackle these, metamorphic testing methods have been employed across different NLP tasks. These methods entail task-specific adjustments at the token or sentence level. For example, in machine translation, this approach might involve replacing a single token in the source sentence to generate variants, whereas in question answering, adjustments might include altering or \nadding sentences within the question or context. By evaluating the system’s responses to these alterations, potential deficiencies in the NLP systems can be identified. Determining the most effective modifications, particularly, especially in terms of which tokens or sentences contribute to system instability, is an essential and continuous aspect of metamorphic testing research. \nTo tackle this challenge, we introduce two white-box methods to detect sensitive tokens in the source text, alterations to which could potentially trigger errors in sequence-to-sequence models. The initial method, termed GRI, leverages GRadient Information for identifying these sensitive tokens, while the second method, WALI, utilizes Word ALignment Information to pinpoint the unstable tokens. We assess these approaches using a Transformer-based model for translation and question answering tasks, comparing them against datasets used by benchmark methods. When applying white-box approaches to \nmachine translation testing and using them to generate test cases, the results show that both GRI and WALI can effectively improve the efficiency of the black-box testing strategies for revealing translation bugs. Specifically, our approaches can always outperform state-of-the-art automatic testing approaches from two aspects: (1) under a certain testing budget (i.e., number of executed test cases), both GRI and WALI can reveal a larger number of bugs than baseline approaches, and (2) when given a predefined testing goal (i.e., number of detected bugs), our approaches always require fewer testing resources (i.e., \na reduced number of test cases to execute). \nAdditionally, we explore the application of GRI and WALI in test prioritization and evaluate their performance in QA software testing. The results show that GRI can effectively prioritize test cases that are highly likely to generate bugs and achieve a higher percentage of fault detection given the same execution budget. WALI, on the other hand, exhibits results similar to baseline approaches, suggesting that while it may not enhance prioritization as significantly as GRI, it maintains a comparable level of effectiveness.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.017 | 0.078 |
| Meta-epidemiology (narrow) | 0.003 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.003 |
| Scholarly communication | 0.003 | 0.008 |
| Open science | 0.004 | 0.004 |
| Research integrity | 0.004 | 0.005 |
| Insufficient payload (model declined to judge) | 0.003 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".