Text summarization based on argumentation techniques

Automatic Text Summarization is one of the Natural Language Process (NLP) studies that aim to build a model for automatically shortening an input text/document while still retaining the necessary information; thus, a conclusion is the output of that model. A machine of Deep Learning (DL) uses a watch-and-learn strategy to imitate task working by consuming a large learning dataset. With the growth of information, DL has become favored and a new basis for most Artificial Intelligence (AI) applications. They have achieved much improved results, even in automatic text summarization. Unfortunately, the DL-based methods are generally constructed under deep and complicated architecture like a block-box, which is uneasy about interpreting.In another corner of AI research, Formal Argumentation aims to develop a formalism system for evaluating arguments based on human-intuitive logical reasoning. This causes Formal Argumentation to be in the spotlight of many tasks that desire transparency. Specifically, Dung's Abstract Argumentation (AA) framework can prove argumentation semantics from arguments, such as acceptability. As abstractive, it does not pay attention to how to construct arguments and attacks by projecting them as nodes and edges in a directed graph. AA is considered a milestone in computational argumentation studies, saying it inspired many researchers to explore internal argumentative structures such as Assumption-based Argumentation (ABA) and ASPIC+.This research aims to mimic how competent readers perform text summarizing by re-constructing the arguments in the text and then arriving at the summary from conclusions of acceptable arguments using formal argumentation techniques. Assuming the availability of the Argumentative Discourse Unit (ADU) graph of the given text, we build structured argumentation frameworks inherited from AA: simplified ASPIC+ and ABA representing the text. Then we use ABA proof procedures to re-construct arguments in the text and evaluate their acceptabilities. In the final stage, we aggregate the conclusions of acceptable arguments.We implement the proof-of-concept experiment of the proposed approach using a small argumentative dataset called Micro-text dataset and compare results to other text summarization techniques. As a result, our summarizer provides comparable results and reasonable relations between text segments of the generated summary. Additionally, we describe the difference between this approach and similar published studies. This research contributes that a transparent AI engine viz. text summarization based on Argumentation. This approach encourages the development of further applications in that their decisions affect humans, such as an automated judicial review or legal case summarization.