Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SYNTHEVAL: Hybrid Behavioral Testing of NLP Models with Synthetic CheckLists

Raoyuan Zhao, Abdullatif Köksal, Yihong Liu, Leonie Weissweiler, Anna Korhonen, Hinrich Schütze

TL;DR

This work proposes SYNTHEVAL, a hybrid behavioral testing framework that leverages large language models (LLMs) to generate a wide range of test types for a comprehensive evaluation of NLP models, and shows that the framework is effective in identifying weaknesses of strong models on these tasks.

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 large language models (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.

SYNTHEVAL: Hybrid Behavioral Testing of NLP Models with Synthetic CheckLists

TL;DR

This work proposes SYNTHEVAL, a hybrid behavioral testing framework that leverages large language models (LLMs) to generate a wide range of test types for a comprehensive evaluation of NLP models, and shows that the framework is effective in identifying weaknesses of strong models on these tasks.

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 large language models (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.
Paper Structure (31 sections, 2 figures, 7 tables)

This paper contains 31 sections, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Synthetic Datasets
  4. Limitations in Handling Linguistic Complexity
  5. Interpretable Behavioral Testing
  6. Methodology: SynthEval
  7. SynthTest: Test Set Generation
  8. SynthTesthard: Challenging Subset
  9. Behavioral Patterns: Formalize and Verify
  10. Analysis
  11. Sentiment Analysis
  12. Negation
  13. Past Tense
  14. Order
  15. Comparative
  16. ...and 16 more sections

Figures (2)

  • Figure 1: Using a held-out val. set for evaluation overestimates the performance. RoBERTaLarge, fine-tuned on 15 diverse sentiment analysis datasets hartmann2023, performs strongly on traditional benchmarks (92.4%). However, SynthEval, which generates behavioral tests with the help of LLMs, demonstrates RoBERTaLarge's bad performance (10.17%) when tested on a sentence containing a simple revision: "I was wrong".
  • Figure 2: A summary of SynthEval with the sentiment analysis task as an example. It consists of three steps: 1. Generating a diverse and large-scale test set with LLMs. 2. Identifying a challenging subset by comparing predictions between a TaskModel (i.e., RoBERTa) and a reference model (i.e., few-shot LLM), and sorting based on differences. 3. Manually designing behavioral patterns and evaluating TaskModels accordingly.