The MARIANNE project-team pursues high-impact research in Artificial Intelligence with a focus on data and models for computational argumentation in natural language. MARIANNE is an Inria joint project-team with the I3S (Computer Science) laboratory of Université Côte d’Azur and CNRS. The team is composed of computer scientists, but it holds a strong interdisciplinary connotation in particular with linguistics, philosophy, sociology and law. MARIANNE proposes innovative Natural Language Processing methods, addressing real-world problems, that are both theoretically sound and explainable. The team focuses on topics such as argument mining, argument and counter-argument generation, argument quality assessment, argument-based explainable AI, argument dynamics, argument-based neuro-symbolic models. The main application scenarios investigated in the team are political debates (propaganda), healthcare, law and online social media (hate speech and disinformation). The team also has a special concern about innovation and transfer. The research project of the team is aligned with the "Core elements of AI" part of the 3IA Côte d’Azur Cluster.

The MARIANNE team aims at developing NLP methods and algorithms for natural language argumentation to address the following three axes.

Axis A - Argument mining:The first research axis will be the development of models and algorithms designed for mining natural language arguments from text. The Argument Mining (AM) scientific community has focused till now on the two main tasks constituting the AM pipeline, i.e., the detection of argumentative components (namely, evidence and claims), and the prediction of the relations of support and attack holding among them. They represent an obligatory starting step, but the resulting argumentation frameworks (i.e., the graph structure composed by the arguments as the nodes of the graph and the relations representing the links) are still quite simple with respect to the needs raised in the team application scenarios. Our goal will be to enhance the extraction of machine-processable natural language argument structures to allow reasoning over complex real world natural arguments, with a focus on the English, French, Italian, Spanish and German languages.

Axis B - Natural language argument quality assessment: The second research axis will be focused on the definition of computational methods to automatically assess the quality of natural language arguments. Despite a few existing approaches, the issue of automatically assessing the quality of an argumentation remains largely unexplored. On the one side, it consists in assessing the quality of the mined arguments to decide, for instance, whether a certain argument has to be selected for synthesising a debate, or whether the overall debate is of good quality. On the other side, it consists in ensuring that the newly generated arguments satisfy the defined quality criteria in order to assess them from the qualitative point of view, i.e., a counter-argument to attack a fake news needs to be concise and without repetitions. The quality of the arguments is also characterized by formal properties of the argumentation graph, e.g., the argument strength, argument preferences, and argument acceptability.

Axis C - Generation of natural language arguments: In addition to the definition of more effective methods to mine (fine-grained) argumentative structures from text, and to automatically assess their quality, the third research axis of the team will consist in the definition of new (generative and not generative) methods to generate natural language arguments, with a focus on English and French initially. This process is incremental and starts with the generation of single argumentative components towards the generation of arguments in the context of interactive dialogues with users. These dialogues are then employed in different use cases with different goals, i.e., explanation, counter-argumentation. These arguments will be firstly generated starting from the mined arguments, and they will rely on abductive reasoning schema based on the set of critical questions and reasoned responses necessary to reach the user's understanding.