Neural Abstractive Summarization for Long Text and Multiple Tables

Abstractive summarization aims to generate a concise summary covering the input document's salient information. Within a report document, the salient information can be scattered in the textual and non-textual content. However, existing document summarization datasets and methods usually focus on the text and filter out the non-textual content. Missing tabular data can limit produced summaries’ informativeness, especially when summaries require covering quantitative descriptions of critical metrics in tables. Existing datasets and methods cannot meet the requirements of summarizing long text and dozens of tables in each report document. To deal with the scarcity of available datasets, we propose FINDSum, the first large-scale dataset for long text and multi-table summarization. Built on 21,125 annual reports from 3,794 companies, FINDSum has two subsets for summarizing each company's results of operations and liquidity. Besides, we present four types of summarization methods to jointly consider text and table content when summarizing reports. Additionally, we propose a set of evaluation metrics to assess the usage of numerical information in produced summaries. Our summarization methods significantly outperform advanced baselines, which verifies the necessity of incorporating textual and tabular data when summarizing report documents. We also conduct extensive comparative experiments to identify vital model components and configurations that can improve summarization results.

I. Introduction

Report documents, like financial reports, investigative reports, and technical reports, are essential information sources. These report documents usually contain large amounts of textual and tabular content and provide rich knowledge about companies, industries, technologies, etc. Each report's salient information can be scattered in long text and multiple tables in different sections, which makes it difficult for non-specialized readers to efficiently read these report documents. A high-quality summary of each report document can help readers quickly browse key information. Automatic document summarization techniques can be utilized to produce reports’ summaries. Users can flexibly adjust the input document and immediately get a summary from the automatic summarization system. Our target is to let the computer generate an informative, fluent, and non-redundant summary for the long text and multiple tables in each report document. To achieve this target, we need to deal with some challenging issues: 1) the scarcity of available datasets, 2) identifying the salient information scattered in a large amount of input content, 3) incorporating different types of content when generating summaries, and 4) models’ efficiency in processing long inputs and outputs.