TL;DR
This paper argues that for Artificial Intelligence to truly integrate with humans in a "hybrid society," machines must be capable of argumentation. The author proposes a research agenda focused on developing AI systems that can understand, generate, and reason with arguments, incorporating socio-cognitive aspects like trust and emotions, and mining arguments from natural language.
What Problem Does This Paper Address?
As AI advances, concerns arise about its societal impact. To foster a "good AI hybrid society," AI systems need to collaborate effectively with humans. This requires AI to move beyond purely logical reasoning and engage in argumentation, mirroring human decision-making processes. Humans naturally use arguments to persuade, evaluate information, and resolve conflicts. Therefore, for AI to participate in mixed human-machine teams, it must be able to handle incomplete and inconsistent information in a way that emulates human argumentation. This paper addresses the challenge of building AI systems that can effectively participate in such argumentative processes.
Methods
The author outlines a three-pillar approach to developing argumentation-enhanced intelligent machines:
- Modeling and Reasoning on Socio-Cognitive Components: This involves developing computational models of argument that can incorporate factors like trust and emotions. The paper discusses fuzzy labeling algorithms for assigning acceptability degrees to arguments based on agent trustworthiness and the use of meta-argumentation to model conflicts about trust.
- Argument Mining in Natural Language: This pillar focuses on extracting argumentative structures from text. It involves NLP techniques to identify argument components (premises, claims) and their relations, as well as KRR for reasoning over these extracted arguments. The paper highlights the emergence of the "Argument(tation) Mining" (AM) field, which aims to analyze discourse pragmatically and apply argumentation theory.
- Correlating Emotions with Arguments: Recognizing that human argumentation is influenced by emotions, this area explores how to integrate emotional aspects into AI models. This includes studying the correlation between emotions, engagement, and mental workload during debates, and assessing the impact of persuasion strategies (Ethos, Pathos, Logos) on human debaters.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Argumentation is Crucial for Human-AI Collaboration: The paper strongly advocates that for AI to be a true partner in a hybrid society, it must be able to argue, not just compute.
- Trust and Emotions are Not Optional: Human decision-making and argumentation are heavily influenced by trust in sources and emotional states. AI systems need to account for these socio-cognitive factors to be effective collaborators.
- Bridging the Gap Between Formal Models and Natural Language: A significant challenge is moving from abstract, formal argumentation models to understanding and generating arguments as expressed in human language. Argument Mining is presented as a key solution.
- The Complexity of Real-World Arguments: Arguments in natural language, especially from social media or debates, are often incomplete, inconsistent, and nuanced, posing a significant challenge for automated analysis.
- Persuasion Strategies Matter: The choice of persuasion strategy (e.g., appealing to emotion vs. logic) can significantly impact mental engagement and the effectiveness of arguments.
Key Results
The paper discusses several research contributions and findings from the author's work and her collaborators:
- Fuzzy Labeling Algorithm: Developed and validated for assigning acceptability degrees to arguments based on agent trustworthiness, showing linear convergence and satisfactory performance on benchmark argumentation graphs.
- Trust as a Multi-Dimensional Concept: Extended fuzzy labeling to formally model trust as a concept relativized to source sincerity and competence, influencing information assessment and providing feedback on source trustworthiness.
- Bipolar Abstract Argumentation: Extended fuzzy labeling to handle support and attack relations, tested on a use case for reconciling information and explaining decisions in a question-answering system.
- Argument Mining on Online Debates: Developed an approach using Textual Entailment and bipolar argumentation to identify arguments and their relations in online debate posts, confirming that support relations are not uniquely interpreted.
- Argument Mining on Social Media: Constructed a dataset (DART) for argument mining on Twitter and explored tasks like fact recognition and source identification, showing feasibility for topics like Grexit and Brexit.
- Argument Mining on Political Speeches: Applied relation prediction techniques to analyze political speeches, supporting researchers in social and political sciences.
- Emotional Impact on Argumentation: Found correlations between emotions, engagement, and mental workload with argumentation elements in human debates. The "Pathos" persuasion strategy was found to be most effective in terms of mental engagement.
Limitations and Future Work
- Multi-disciplinarity Challenges: The research requires significant collaboration with experts in linguistics, cognitive science, and psychology, necessitating the development of a shared vocabulary.
- Incomplete Representation of Socio-Cognitive Components: Formalizing fuzzy, multi-faceted elements like trust and emotions can lead to incomplete representations, with the risk of over-specializing in one aspect while neglecting others.
- Complexity of Natural Language Arguments: Arguments in the wild (e.g., social media) are far from prototypical patterns, increasing the difficulty of argument mining tasks.
- Performance on Diverse Documents: While good results have been achieved in some scenarios, system performance needs improvement for other document types, such as legal cases.
- Human Agreement as a Benchmark: Even human agreement on annotation tasks is affected by the complexity of argument mining tasks, indicating the inherent difficulty of the field.
Future directions include continuing to make artificial argumentation closer to human argumentation, focusing on dealing with incomplete, conflicting, and uncertain information, and further developing computational models of argument and argument mining techniques.
Summary updated