A Field Guide to Automatic Evaluation of LLM-Generated Summaries

Abstract

Large Language models (LLMs) are rapidly being adopted for tasks such as text summarization, in a wide range of industries. This has driven the need for scalable, automatic, reliable, and cost-effective methods to evaluate the quality of LLM-generated text. What is meant by evaluating an LLM is not yet well defined and there are widely different expectations about what kind of information evaluation will produce. Evaluation methods that were developed for traditional Natural Language Processing (NLP) tasks (before the rise of LLMs) remain applicable but are not sufficient for capturing high-level semantic qualities of summaries. Emerging evaluation methods that use LLMs to evaluate LLM-output, appear to be powerful but lacking in reliability. New elements of LLM generated text that were not an element of previous NLP tasks, such as the artifacts of hallucination, need to be considered. We outline the different types of LLM evaluation currently used in the literature but focus on offline, system-level evaluation of the text generated by LLMs. Evaluating LLM-generated summaries is a complex and fast-evolving area, and we propose strategies for applying evaluation methods to avoid common pitfalls. Despite having promising strategies for evaluating LLM summaries, we highlight some open challenges that remain.