SYNTHEVAL: Hybrid Behavioral Testing of NLP Models with Synthetic CheckLists (2408.17437v2)

Abstract: Traditional benchmarking in NLP typically involves using static held-out test sets. However, this approach often results in an overestimation of performance and lacks the ability to offer comprehensive, interpretable, and dynamic assessments of NLP models. Recently, works like DynaBench (Kiela et al., 2021) and CheckList (Ribeiro et al., 2020) have addressed these limitations through behavioral testing of NLP models with test types generated by a multistep human-annotated pipeline. Unfortunately, manually creating a variety of test types requires much human labor, often at prohibitive cost. In this work, we propose SYNTHEVAL, a hybrid behavioral testing framework that leverages LLMs to generate a wide range of test types for a comprehensive evaluation of NLP models. SYNTHEVAL first generates sentences via LLMs using controlled generation, and then identifies challenging examples by comparing the predictions made by LLMs with task-specific NLP models. In the last stage, human experts investigate the challenging examples, manually design templates, and identify the types of failures the taskspecific models consistently exhibit. We apply SYNTHEVAL to two classification tasks, sentiment analysis and toxic language detection, and show that our framework is effective in identifying weaknesses of strong models on these tasks. We share our code in https://github.com/Loreley99/SynthEval_CheckList.

• The paper introduces a novel hybrid approach that integrates synthetic data generation and manual pattern verification to reveal hidden flaws in NLP models.
• The methodology uses output probability differences to pinpoint challenging test cases, enabling detailed error analysis in tasks like sentiment analysis and toxic language detection.
• Results demonstrate that conventional benchmarks can overlook specific vulnerabilities, emphasizing the need for more robust and scalable evaluation frameworks in NLP.

SynthEval: Hybrid Behavioral Testing of NLP Models with Synthetic CheckLists

SynthEval introduces a methodology aimed at advancing the evaluation of NLP models through a hybrid approach that leverages both synthetically generated test sets and human annotations. This paper challenges the predominant reliance on traditional held-out test sets, demonstrating the obliviousness of such tests to nuanced performance flaws in NLP models. SynthEval provides an efficient, scalable alternative that uncovers both broad and granular weaknesses in NLP models, allowing for a more rigorous performance appraisal.

Overview of the SynthEval Framework

The SynthEval framework is partitioned into three distinct stages:

1. Diverse Synthetic Test Set Generation (SynthTest): This initial stage entails using LLMs, specifically the LLaMA2 7B model, to generate a diverse array of sentences for the task at hand, inspired by a randomly sampled set of seed words from an existing dataset.
2. Identification of Challenging Test Subsets (SynthTest\textsubscript{hard}): The generated sentences are subjected to both the target task-specific models (TaskModels) and the reference LLM in order to identify discordance in their predictions. By sorting these sentences based on the absolute difference in output probabilities, a subset of challenging examples is determined.
3. Manual Formalization and Verification of Behavioral Patterns: Annotators manually inspect the difficult examples to identify recurring patterns of failure, which are then formalized into behavioral templates. These templates allow for the automated generation of numerous similar challenging examples, providing a robust mechanism for stress-testing NLP models.

Results and Analysis

The paper examines two NLP tasks: sentiment analysis and toxic language detection. The results for each task are categorized into several failure patterns, offering insights into the nuanced weaknesses of models like RoBERTa, DistilBERT, and ToxDetect.

Sentiment Analysis

For sentiment analysis, two task-specific models—SiEBERT (a variant of RoBERTa) and DistilBERT—were evaluated. Despite impressive overall results on standard benchmarks, these models exhibited substantial errors when confronted with sentences containing complex linguistic structures such as negation and revisions. Examples of these include:

• Negation: Sentences like "I don't think this movie is awful." led to drastically reduced accuracies for DistilBERT and RoBERTa, showing their difficulty in handling linguistic negations.
• Past Tense Revisions: Patterns where initial opinions are revised, such as "I thought this movie was great. I was wrong," also caused significant drops in performance, highlighting issues with context comprehension.
• Order and Specific Phrases: The performance discrepancies in understanding idiomatic expressions or changes in word order pointed to incomplete linguistic models under different grammatical constructs.

Toxic Language Detection

For toxic language detection, ToxDetect and DistilBERT were assessed using sentences generated via similar principles. Here, the models displayed specific vulnerabilities such as:

• Nonsense Characters: Adding random non-alphabetic characters significantly degraded model performance, with neither model handling these gracefully.
• Ethnic Slurs: A prominent weakness was exposed in the models' capacity to correctly identify various ethnic slurs, showing a high rate of false negatives.
• Negatively Sentimented Non-toxic Statements: Sentences with a negative sentiment were misclassified as toxic, reflecting an over-generalization problem in distinguishing between negativity and toxicity.

Implications for Future AI Developments

SynthEval's ability to automatically generate diverse, nuanced behavioral tests elucidates specific shortcomings in NLP models that are obfuscated by traditional evaluation methods. This capability presents several implications:

1. Richer Model Validation: The increased granularity in test cases allows for a more exacting evaluation of model performance, identifying edge cases and specific areas of deficiency that can be overlooked by simpler aggregate metrics.
2. Model Robustness: Identifying and addressing the kinds of hidden vulnerabilities exposed by SynthEval can contribute to training more resilient NLP models, better equipped to handle the rich complexity and variability of human language.
3. Template-based Behavioral Patterns: The formalization of failure patterns into templates facilitates systematic stress-testing across various models and languages, paving the way for more adaptive and scalable evaluation frameworks.

Conclusion

SynthEval demonstrates the efficacy of hybrid behavioral testing that blends automation with human insights. This approach not only substantiates existing high accuracies but unearths previously concealed weaknesses of NLP models. Furthermore, it offers a pathway towards developing more robust, versatile, and adaptable NLP systems, highlighting both immediate improvements and long-term advancements in the field of artificial intelligence. Future work may extend SynthEval to more complex, multi-class tasks and incorporate additional layers of analysis to enhance the identification and formalization of intricate language patterns, such as irony and figurative speech.