# Long Text and Multi-Table Summarization: Dataset and Method

**Authors:** Shuaiqi Liu, Jiannong Cao, Ruosong Yang, Zhiyuan Wen  
**arXiv:** [2302.03815](https://arxiv.org/abs/2302.03815)  
**Date:** 2023-02-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.03815v1/

## Key points

- Automatic document summarization is important in natural language processing.
- Existing methods and datasets focus on text and filter out non-textual content, limiting informativeness.
- Missing tabular data can be a significant issue for quantitative descriptions of critical metrics in tables.
- Shuaiqi Liu et al. propose FINDSum - the first large-scale dataset for long text and multi-table summarization.
- The dataset is built on 21,125 annual reports from 3,794 companies and has two subsets for summarizing each company's results of operations and liquidity.
- Three types of summarization methods are presented: extractive-based, abstractive-based, and hybrid-based.
- A set of evaluation metrics is proposed to assess the usage of numerical information in produced summaries accurately.
- Jointly considering input textual and tabular data is crucial when summarizing report documents.
- This work provides an essential contribution to advancing automatic document summarization by addressing some significant limitations in current approaches while also providing a new benchmark dataset for future research efforts.

## Summary

In the field of natural language processing, automatic document summarization is a crucial task that aims to produce a concise summary covering the input document's salient information. However, existing document summarization datasets and methods usually focus on the text and filter out non-textual content, which can limit produced summaries' informativeness. In particular, missing tabular data can be a significant issue when summaries require covering quantitative descriptions of critical metrics in tables. To address these limitations, Shuaiqi Liu et al. propose FINDSum - the first large-scale dataset for long text and multi-table summarization. This dataset is built on 21,125 annual reports from 3,794 companies and has two subsets for summarizing each company's results of operations and liquidity. To summarize the long text and dozens of tables in each report effectively, they present three types of summarization methods: extractive-based methods that select important sentences from the input text; abstractive-based methods that generate new sentences based on the input; and hybrid-based methods that combine both extractive and abstractive approaches. Furthermore, to assess the usage of numerical information in produced summaries accurately, they propose a set of evaluation metrics. Dataset analyses and experimental results indicate that jointly considering input textual and tabular data is crucial when summarizing report documents. Overall, this work provides an essential contribution to advancing automatic document summarization by addressing some significant limitations in current approaches while also providing a new benchmark dataset for future research efforts.