A Novel Computational Workflow for Argumentation in AI

In this project we investigate ways of performing automated reasoning based on argument structures. Within the area of Artificial Intelligence, our field of study is situated in Computational Argumentation. In this field one studies how to represent arguments, what are relationships between arguments, and how to automatically calculate reasoning, based on such arguments. Arguments, as we see or hear them in everyday life, have vary different forms, structures, and aims: for instance, they can be based on authority (arguing by ones expertise), they can connect emotionally (appeal to emotions), or argue on rational grounds (using, e.g., logical reasoning), to name some examples. Structurally, they can be simple or very complex. For our aim of performing automated reasoning on argument structures, we rely on formal specifications of arguments and their inter-relationship (e.g., arguments countering or supporting one another). Such formal specifications allow for implementation of a program that computes argumentative outcomes. In Computational Argumentation, many ways of defining structures of arguments have been proposed, to capture various aspects of argumentation. Importantly, the most common approaches here assume that one constructs arguments and their relationships from given knowledge, and one computes argumentative outcomes after having constructed these arguments. That is, one tries to formulate all ways in which one can argue, based on ones knowledge. While this is a conceptually clean way, in general it can be the case that many arguments are constructed in this way. Having many arguments can be a barrier to computation (the computer has to consider many arguments), and a large number of arguments can be a barrier to human understanding. In our project, we study ways of performing argumentative reasoning differently: in contrast to many earlier approaches, we aim to calculate argumentative reasoning, at first, without making arguments explicit. While this is challenging without access to the concrete arguments, preliminary research shows that with carefully looking at ones knowledge one can reason without explicated arguments. We then, after having concluded what should be the argumentative outcome, plan to construct only arguments required to show the result. In this way, we aim for a novel workflow: compute the result on the original knowledge base and present arguments only when required. For this to be realized, we need to understand argumentative reasoning at a deep level, which is a major challenge for our research project. Based on this, we aim to provide academic prototypes of our approach, both for showing strengths and potential weaknesses of our approaches, and for others to experiment on. We think that our research project can support Computational Argumentation by providing novel insights into computation of argumentative outcomes, and new academic tools.

In this research project we studied computational argumentation. Argumentation is a long-standing field, dating back to classical Greek philosophy. The "computational" variant, dubbed computational argumentation, is a subfield of modern Artificial Intelligence (AI), in which the research agenda is on how to represent arguments, how to relate different arguments, and how to perform reasoning on such arguments. For instance, an argument might consist of premises and a conclusion, like "Jan claims that John is in Amsterdam", with the conclusion being that John is in Amsterdam. Another argument might conclude that Jan is not truthful about John's location, leading to a counter-argument to the former argument. Based on such arguments and their relation, one can define how argumentative reasoning is carried out, e.g., one can find an argument acceptable if there is no reason to the contrary, i.e., no counter-argument known. If a counter-argument becomes available, the status depends on the counter-argument. In this research project we looked at a computational "workflow" of how to carry out argumentative reasoning as outlined above. In general, one faces challenges like complex argument structures, many arguments, and possibly complicated relations between arguments. We addressed such challenges by combining several aspects: we make use of internal structure of arguments, information about strength of arguments (is one stronger than another?), degree of belief in arguments, and properties of argumentative reasoning. We designed novel algorithmic approaches and open-access software tools that perform argumentative reasoning efficiently, and tested these tools in various ways. For instance, we organized an international competition of such tools, comparing state-of-the-art solutions from around the world, and we collaborated with the Dutch National Police on an application for argumentative reasoning. We extended these findings also with approaches how to simplify the presentation of results of argumentative reasoning. We devised methods of presenting complex argument networks and complex internal structure by abstracting possibly irrelevant parts away, but keeping the argumentative reasoning sound. Overall, this research project advanced several aspects of automated argumentative reasoning, yet, several challenges remain for future work. In particular, we think that (i) ways of visualizing and presenting simplified argumentative reasoning to advance understanding and (ii) connecting state-of-the-art machine learning approaches with computational argumentation, e.g., in the form of learning how to construct arguments or their internal structure, can lead to further advancement in wider applicability of computational argumentation.