FactorSum is an abstractive summarization model that generates snippets with relevant information from the input document while taking into account user-defined guidance such as a desired number of words or similarity to a given textual content. We found that separating (factorizing) information relevance judgements from user preferences improves the adaptability of summarizers across different domains.
See here for our demo.
In this demo, you can inspect generated summaries for documents from arXiv, PubMed, and GovReport. You can also tweak some parameters such as budget guidance and perform cross-domain predictions. For more information, refer to the code repository.
The paper is here.
To read more about FactorSum, see the paper that can be downloaded here.
author = "Fonseca, Marcio and Ziser, Yftah and Cohen, Shay B.",
title = "Factorizing Content and Budget Decisions in Abstractive Summarization of Long Documents",
booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
year = "2022",
publisher = "Association for Computational Linguistics",
location = "Abu Dhabi"