Evaluating Text Summaries Generated by Large Language Models Using OpenAI's GPT

TL;DR

Problem: evaluating summaries from transformer models in data-rich environments. Approach: use GPT-3.5 as an independent evaluator alongside traditional metrics (ROUGE, LSA, Flesch-Kincaid) on CNN/Daily Mail summaries generated by six Hugging Face models; employ varied prompts (zero-shot, chain-of-thought, scoring) to assess four properties: conciseness, relevance, coherence, readability. Findings: GPT-based evaluations correlate strongly with relevance and coherence from traditional metrics and tend to yield higher scores; ProphetNet is notably limited by conciseness. Significance: demonstrates that AI-driven evaluation can complement traditional metrics for more nuanced, robust assessment of summarization systems, informing model selection and benchmarking in NLP tasks.

Abstract

This research examines the effectiveness of OpenAI's GPT models as independent evaluators of text summaries generated by six transformer-based models from Hugging Face: DistilBART, BERT, ProphetNet, T5, BART, and PEGASUS. We evaluated these summaries based on essential properties of high-quality summary - conciseness, relevance, coherence, and readability - using traditional metrics such as ROUGE and Latent Semantic Analysis (LSA). Uniquely, we also employed GPT not as a summarizer but as an evaluator, allowing it to independently assess summary quality without predefined metrics. Our analysis revealed significant correlations between GPT evaluations and traditional metrics, particularly in assessing relevance and coherence. The results demonstrate GPT's potential as a robust tool for evaluating text summaries, offering insights that complement established metrics and providing a basis for comparative analysis of transformer-based models in natural language processing tasks.

Figures (5)

• Figure 1: Generation of summaries.
• Figure 2: Illustration of the GPT-Based evaluation.
• Figure 3: Illustration of the Metrics-Based evaluation.
• Figure 4: Assessing using traditional metrics.
• Figure 5: Assessing using GPT.