SummEval: Re-evaluating Summarization Evaluation

TL;DR

This paper critically re-evaluates 14 automatic summarization metrics, benchmarks 23 models, and provides a comprehensive dataset and toolkit to improve evaluation practices and align them more closely with human judgments.

Contribution

It offers a thorough re-assessment of evaluation metrics, a large dataset of model summaries and human judgments, and an extensible toolkit to advance summarization evaluation research.

Findings

14 metrics re-evaluated comprehensively

Benchmarking of 23 recent models conducted

Largest collection of human judgments assembled

Abstract

The scarcity of comprehensive up-to-date studies on evaluation metrics for text summarization and the lack of consensus regarding evaluation protocols continue to inhibit progress. We address the existing shortcomings of summarization evaluation methods along five dimensions: 1) we re-evaluate 14 automatic evaluation metrics in a comprehensive and consistent fashion using neural summarization model outputs along with expert and crowd-sourced human annotations, 2) we consistently benchmark 23 recent summarization models using the aforementioned automatic evaluation metrics, 3) we assemble the largest collection of summaries generated by models trained on the CNN/DailyMail news dataset and share it in a unified format, 4) we implement and share a toolkit that provides an extensible and unified API for evaluating summarization models across a broad range of automatic metrics, 5) we assemble and share the largest and most diverse, in terms of model types, collection of human judgments of model-generated summaries on the CNN/Daily Mail dataset annotated by both expert judges and crowd-source workers. We hope that this work will help promote a more complete evaluation protocol for text summarization as well as advance research in developing evaluation metrics that better correlate with human judgments.