Toward Computational Argumentation with Reasoning and Knowledge

 Our society today is overloaded with information and opinions. While they are important resources for decision-making for the general public and policy makers in organizations, the staggering amount of them is making people more passive and dependent on information delivered by technologies. This issues an urgent call for technologies that support human decision-making in a truthful way. Truthful language technologies need the ability to reason and use knowledge, beyond memorizing patterns in data and relying on irrelevant statistics and biases. To achieve this goal, our field needs a better understanding of how humans reason and how to incorporate human-like reasoning and knowledge into computational models. 

In response to this need, this thesis studies one of the most common communication modes that is full of reasoning: argumentation. The first goal is to provide computational models for analyzing argumentation quantitatively and shedding light on human reasoning reflected in language. The second goal is to incorporate the findings from our study and argumentation theory into computational models via proper knowledge to improve their predictive power and fidelity. By doing so, this thesis argues that integrating reasoning and knowledge, along with argumentation theory, into computational models improves their explanatory and predictive power for argumentative phenomena. 

This thesis begins with a study of individual statements in argumentation, in terms of asserted propositions, propositional types, and their effects. We build a model that identifies argumentatively meaningful text spans in text and recovers asserted propositions. Next, we present a methodology for identifying various surface types of propositions (e.g., statistics and comparison) that underlie dialogues and analyzing their associations with different argumentation outcomes (e.g., persuasion). Applying the model on four argumentative corpora, we find 24 generic surface types of propositions in argumentation and their associations with successful editing in Wikipedia, moderation in political debates, persuasion, and formation of pro- and counter-arguments. 

We take a step further and study argumentative relations between statements (support, attack, and neutral) by drawing upon argumentation schemes. We first address the challenging problem of annotation in application of argumentation schemes to computational linguistics. We develop a human-machine hybrid annotation protocol to improve the speed and robustness of annotation. By applying it to annotating four main types of statements in argumentation schemes, we demonstrate the natural affinity between the statement types to form arguments and argumentation schemes. Next, we hypothesize four logical mechanisms in argumentative relations informed by argumentation theory: factual consistency, sentiment coherence, causal relation, and normative relation. Not only do they explain argumentative relations effectively, but incorporating them into a supervised classifier through representation learning further improves the predictive power by exploiting intuitive correlations between argumentative relations and logical relations. 

Lastly, we take a closer look at counter-argumentation and study counterargument generation. We first present two computational models to detect attackable sentences in arguments via persuasion outcomes as guidance. Modeling sentence attackability improves prediction of persuasion outcomes. Further, they reveal interesting and counterintuitive characteristics of attackable sentences. Next, given statements to attack, we build a system to retrieve counterevidence from various sources on the Web. At the core of this system is a natural language inference (NLI) model that classifies whether a candidate sentence is valid counterevidence to the given statement. To overcome the lack of reasoning abilities in most NLI models, we present a knowledge-enhanced NLI model that targets causality- and example-based inference. This NLI model improves performance in NLI tasks, especially for instances that require the targeted inference, as well as the overall retrieval system. We conclude by making a connection of this system with the argumentative relation classifier and attackability detection. 

The contributions of the thesis include the following: 

• This thesis contributes computational tools and findings to the growing literature of argumentation theory on quantitative understanding of argumentation. 
• This thesis provides insights into human reasoning and incorporates them into computational models. For instance, logical mechanisms are incorporated into an argumentative relation classifier, and two types of inference are incorporated into counterevidence retrieval through relevant knowledge graphs. 
•  This thesis draws largely on and borrows frameworks from argumentation theory, thereby bridging argumentation theory, language technologies, and computational linguistics.