The Development of Single-Document Abstractive Text Summarizer During the Last Decade

Introduction

The digital world has become more complex and crowded with massive volumes of digital data. In 2018, the size of the indexed World Wide Web is over 5.22 billion pages (Kunder, 2018, Dec 15), spread over 1.8 billion websites (Fowler, 2018, Feb 20). As the number of electronic text documents is growing so is the need for an automatic text summarizer. It is difficult to read and gain knowledge from vast number of texts. In some fields, reading and understanding long texts consume time and effort. Consequently, automatic text summarization can be seen as a viable solution which is used in different domains and applications. It can decrease the time taken to summarize huge texts in many areas and media. It extracts and identifies the important information from a text; which can provide concise information with less effort and time. In addition, it can solve the information storage problem by reducing the document’s size. Text summarization can support different applications and usage such as news feed, reports abstract, meeting, email and email threads, digest web pages and blogs, recap large amount of web opinions, helping doctors to get an overview about their patients’ medical history. Also, students can use the summarization as a helping tool for quick overviewing their studying materials. Web crawler bots can be used to browse the web systematically according to a specific field; news for example, and summarize their contents in a meaningful way. Text summarization can be used in various stand-alone applications or combined with other systems, such as information retrieval, text clustering, data mining applications, web documents and pages, tweet, and opinion summarization.

Automatic text summarization is not a new idea, but there is a huge room for improvement. Simulating how human summaries texts lead to a major innovation in the field of artificial intelligence, abstractive text summarization becomes a necessity in the field of Natural Language Processing. It needs multiple tools to run together in order to extract knowledge and generate a new text. Many researchers have focused on extractive method due to its simplicity. Even though the extractive summarization is quite advanced, there are still researchers working on single document summarization, and those working on multi-document summarization. Now, the abstractive summarization itself is a challenging area. There are few research studies about abstractive summarization in different languages which are still immature due to the difficulties and the challenges of the natural language generation. More effort is necessary to advance this important research field.

In the literature, abstractive text summarization has been applied on several languages; such as English, Arabic, Hindi, Kannada, Malayalam, Telugu, and Vietnamese. Different methods have been conducted to achieve abstractive summary; such as discourse structure, graph-base, semantic-base, linguistic-based, information extraction rules, statistical model, machine learning techniques which include deep learning methods, and hybrid methods. Some of abstractive text summarization systems use the output of extractive summarization systems as a middleware to produce an abstractive summary such as (Kallimani, Srinivasa, & Eswara Reddy, 2011) and (Lloret, Romá-Ferri, & Palomar, 2013). Another troubling issue is the lack of automated assessment tool. For extractive summaries, researchers have developed different metrics to automatically assess the quality of the summary. However, for abstractive summaries, it is mainly human judgment. This is an expensive and time-consuming method for assessing the summary, and probably one of the main contributor to slow pace of development of abstractive summarizers.

This chapter will focus on the state-of-the-art work on abstractive text summarization systems for single documents, regardless of the target language of the input document, with a research has been published in English. To give room for in-depth look, the authors decided to limit the review to those developed during the last decade (covering the period 2008-2018), exploring the methods used in this subject and identifying their limitations. Furthermore, the chapter will show how these systems are evaluated and which datasets have been used. The fact is, abstractive text summarization is much more challenging than its extractive counterpart, and as such there are a few works in this area in all the languages.