# SAMSum Corpus: A Human-annotated Dialogue Dataset for Abstractive Summarization

**Authors:** Bogdan Gliwa, Iwona Mochol, Maciej Biesek, Aleksander Wawer  
**arXiv:** [1911.12237](https://arxiv.org/abs/1911.12237)  
**Date:** 2019-11-27  
**Summary page:** https://www.summarizepaper.com/en/paper/1911.12237v2/

## Key points

- Introduction of SAMSum Corpus, a new dataset for abstractive dialogue summarization
- Investigation of challenges in abstractive dialogue summarization through testing various models
- Comparison of model performance with news articles corpus
- Surprising finding: model-generated summaries of dialogues had higher ROUGE scores than news articles, despite human evaluators' judgement suggesting otherwise
- Importance of dedicated models and non-standard quality measures in abstractive dialogue summarization
- SAMSum Corpus: high-quality collection of chat-dialogues manually annotated with abstractive summaries, valuable resource for further research.

## Summary

This paper introduces the SAMSum Corpus, a new dataset specifically designed for abstractive dialogue summarization. The authors investigated the challenges posed by this task by testing various models and comparing their performance with that on a corpus of news articles. Surprisingly, the model-generated summaries of dialogues achieved higher ROUGE scores compared to those of news articles, despite human evaluators' judgement suggesting otherwise. This finding highlights the need for dedicated models and non-standard quality measures in abstractive dialogue summarization. The SAMSum Corpus is a high-quality collection of chat-dialogues manually annotated with abstractive summaries, making it a valuable resource for further research in this field.