SynthTextEval: Synthetic Text Data Generation and Evaluation for High-Stakes Domains

Krithika Ramesh, Daniel Smolyak, Zihao Zhao, Nupoor Gandhi, Ritu Agarwal, Margrét V. Bjarnadóttir, Anjalie FieldEMNLP 2025

TL;DR

This paper introduces SynthTextEval, a comprehensive toolkit for evaluating synthetic text data, crucial for privacy-preserving AI development in sensitive domains like healthcare and law. It addresses the lack of standardized evaluation methods by offering a unified framework for assessing utility, privacy, fairness, and quality, enabling more robust and trustworthy AI systems.

What Problem Does This Paper Address?

Developing AI systems for high-stakes domains like healthcare and law often requires access to sensitive, private data. While synthetic data generation offers a promising avenue for privacy preservation, its adoption is hampered by a lack of standardized and comprehensive evaluation methods. Existing approaches often focus on isolated aspects of synthetic data quality (e.g., utility or privacy) and use varying metrics, making it difficult to compare different generation methods or ensure that synthetic data is truly suitable for its intended purpose. This inconsistency can lead to a false sense of security regarding privacy or an underestimation of synthetic data's utility, ultimately hindering the development of reliable and ethical AI in critical fields.

Methods

The core contribution is the SynthTextEval toolkit, which provides a modular approach to evaluating synthetic text. Its key features include:

  • Text Generation Module: Capable of generating synthetic text with optional differential privacy (DP) guarantees using DP-SGD. This module allows for fine-tuning LLMs with control codes to guide generation based on desired characteristics.
  • Evaluation Modules: A suite of generalizable metrics covering:
  • Downstream Utility: Assesses how well synthetic data can train models for tasks like classification and coreference resolution. It uses "silver" annotations generated by pre-trained models for efficiency.
  • Fairness: Evaluates models trained on synthetic data using metrics like Equalized Odds (EO) and False Positive Equality Difference (FPED) to ensure equitable performance across subgroups.
  • Privacy: Employs canary-based evaluations to detect memorization and entity-centric metrics to quantify the leakage of sensitive information and personal identifiers (PII).
  • Text Quality: Utilizes automated metrics like Perplexity, Fréchet Inception Distance (FID), and MAUVE to assess the fluency and distributional similarity of synthetic text to real data.
  • User-Friendly GUI: A visual interface for comparing synthetic and real text, highlighting matching entities, and allowing domain experts to provide qualitative feedback. This is crucial for capturing nuances not easily captured by automated metrics.

Datasets

The paper uses the following datasets for evaluation:

  • Text Anonymization Benchmark (TAB): A dataset of 1,268 European court cases, used for evaluating synthetic data generation for privacy preservation and downstream classification tasks.
  • MIMIC-III: A critical care database, used for evaluating synthetic data generation for downstream tasks like ICD-9 code prediction and coreference resolution.
  • i2b2/VA Shared-Task: Coreference-annotated clinical notes, also used for evaluating coreference resolution.

The authors do not introduce a new benchmark dataset but rather demonstrate the utility of their toolkit on existing, relevant datasets from high-stakes domains.

Key Insights and Ideas

  • The Trade-off is Real: The study consistently shows a trade-off between differential privacy and utility/quality. While DP synthetic data offers better privacy, it generally leads to lower performance in downstream tasks and reduced text quality compared to non-DP synthetic data or real data.
  • Context Matters for Privacy: Simple PII leakage metrics are not enough. The paper highlights the importance of considering the context in which entities appear, as demonstrated by the canary attack results showing increased privacy risk with longer context windows.
  • Qualitative Feedback is Essential: Automated metrics can miss subtle issues. The GUI's ability to solicit feedback from non-technical domain experts is crucial for understanding the practical implications of synthetic data quality and privacy in specific contexts (e.g., a judge's name leakage might be acceptable in law, but a patient's name in healthcare is not).
  • Standardization is Key: The lack of standardized evaluation frameworks is a significant barrier. SynthTextEval aims to bridge this gap by providing a unified set of metrics and tools, fostering comparability across research.

Key Results

  • TAB Dataset:
    • Synthetic data was roughly half the length of real data.
  • Non-DP synthetic data (Dε=∞) showed higher Jaccard and Cosine similarity to real data compared to DP synthetic data (Dε=8).
  • FID scores were comparable across all datasets, but MAUVE and perplexity indicated lower quality for synthetic data.
  • Models trained on non-DP synthetic data achieved higher F1 and accuracy than those trained on DP synthetic data.
  • DP synthetic data (Dε=8) had a lower percentage of leaked PII compared to non-DP synthetic data.
  • Healthcare (MIMIC-III) Dataset:
    • Training on either synthetic dataset resulted in significantly lower utility for coreference resolution and mention detection compared to real data.
  • For ICD-9 classification, DP synthetic data (Dε=8) showed a more pronounced drop in utility than non-DP synthetic data (Dε=∞), though both lagged behind real data.
  • Fairness metrics indicated that models trained on real data were generally the most fair, with DP synthetic data often showing poorer fairness for race as a sensitive attribute.
  • DP synthetic data (Dε=8) included a lower percentage of reproduced entities compared to non-DP synthetic data (Dε=∞).

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the performance of synthetic data is influenced by hyperparameter choices during generation.
  • The TAB dataset, being small, led to relatively high privacy risks for both synthetic datasets.
  • While DP provides theoretical privacy guarantees, the paper notes that achieving both high privacy and high utility simultaneously remains a challenge.
  • Future Work:
    • The authors are committed to improving the toolkit by widening the evaluation criteria and metrics.
  • They plan to provide further visual or interactive tools for comparison.
  • Enhancing the reliability of the package is also a stated goal.
  • Future efforts will focus on making the toolkit more adaptable to various contexts and use cases.

Summary updated

BibTeX
@inproceedings{ramesh2025synthtexteval,
  author = {Krithika Ramesh and Daniel Smolyak and Zihao Zhao and Nupoor Gandhi and Ritu Agarwal and Margrét V. Bjarnadóttir and Anjalie Field},
  title = {{SynthTextEval: Synthetic Text Data Generation and Evaluation for High-Stakes Domains}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure