DEE: Dual-Stage Explainable Evaluation Method for Text Generation

Shenyu Zhang, Yu Li, Rui Wu, Xiutian Huang, Yongrui Chen, Wenhao Xu, Guilin Qi

Introduction

Evaluating Large Language Models (LLMs) presents a significant challenge in industrial settings. While traditional metrics like ROUGE or BERTScore are computationally efficient, they lack the depth to explain why a generated text is flawed. Conversely, modern LLM-based evaluators like GPT-4 provide sophisticated, explainable feedback but are often too slow and costly for real-time monitoring.

To bridge this gap, researchers have introduced DEE (Dual-stage Explainable Evaluation). This framework balances the immediate needs of online text monitoring with the granular diagnostic requirements of model debugging, outperforming GPT-4 in human judgment correlation while maintaining industrial-grade efficiency.

The DEE Framework

DEE is built upon a Llama-2-7B backbone and utilizes a novel dual-stage architecture. This design allows the system to serve two distinct operational needs through stage-specific instructions:

  1. Stage I: Fast Detection (Online Scenarios): This stage is optimized for speed. It quickly determines if a text contains errors and classifies them into broad categories such as "Reliability" or "Toxicity." By outputting a minimal number of tokens, it functions as a high-speed filter suitable for real-time deployment.
  2. Stage II: Detailed Diagnosis (Offline Scenarios): For deeper analysis, this stage generates a comprehensive diagnostic report. It identifies the exact location of the error, provides a natural language explanation, assigns a severity score, and calculates a final quality metric.
The DEE framework: Stage I for fast online detection and Stage II for detailed offline diagnosis.

The AntEval Dataset

To train and validate DEE, the researchers developed AntEval, a dataset comprising 15,000 examples derived from real-world applications at Alipay. Unlike many academic benchmarks, AntEval focuses on issues encountered in live production environments, categorized into three core dimensions:

  • Reliability: Addressing hallucinations and instruction misalignment.
  • Bias & Toxicity: Identifying social biases and offensive content.
  • Basic Dimensions: Evaluating fundamental fluency and coherence.

AntEval spans four diverse tasks: Social Media Post Generation (SMPG), Dialogue Generation (DG), Text Paraphrase (TP), and Story Generation (SG). This variety ensures the evaluator is robust across different linguistic contexts.

Distribution of error categories and task-specific error counts in the AntEval dataset.

Key Results and Insights

Experimental results demonstrate that DEE is not only faster but also more accurate than existing solutions:

  • Superior Human Correlation: DEE achieved a Pearson correlation of 53.7%, surpassing GPT-4 (48.1%) and specialized metrics like TIGERScore (37.1%).
  • Decomposition Advantage: The researchers found that splitting the evaluation into two stages actually improves accuracy. By first identifying the error category in Stage I, the model is less prone to hallucinations when generating the detailed explanation in Stage II.
  • High Veridicality: Expert review confirmed that DEE’s diagnostic reports are highly reliable, with a "Veridicality Rate" (the precision of identified errors) between 89% and 93%.
  • Operational Efficiency: Stage I operates at speeds comparable to lightweight models like T5-base, making it a viable solution for high-traffic production pipelines.

Conclusion

DEE represents a significant step forward in making LLM evaluation both practical and explainable. By leveraging a dual-stage approach and the real-world grounding of the AntEval dataset, it provides a scalable solution for maintaining text quality in industrial applications. While future work will focus on handling extremely "noisy" texts with multiple overlapping errors, DEE currently offers a state-of-the-art balance of speed and depth.

Summary updated

BibTeX
@inproceedings{zhang2024dee,
  author = {Shenyu Zhang and Yu Li and Rui Wu and Xiutian Huang and Yongrui Chen and Wenhao Xu and Guilin Qi},
  title = {{DEE: Dual-Stage Explainable Evaluation Method for Text Generation}},
  booktitle = {DASFAA},
  year = {2024},
  doi = {10.1007/978-981-97-5575-2\_29},
}

Loading related content...

Figure