WP1 – Argument Structure Parsing
WP1 develops methods to automatically identify and extract arguments, counter-arguments, and preferences from natural language text. It focuses on training robust argument parsers using machine learning, adapting to multiple domains such as legal or medical texts, and tackling data scarcity through multi-task and weak supervision techniques.
WP2 – Synthetic Data Generation for Argument Parsers
This WP generates large, high-quality datasets by producing artificial argumentative texts and structures. Using back-translation and generative AI models, WP2 supports WP1 by providing diverse and realistic training data to improve parser accuracy and generalization.
WP3 – Knowledge Revision
WP3 designs mechanisms allowing GRAIL agents to revise their beliefs and arguments when confronted with new or contradictory information. The aim is to maintain internal consistency and enable agents to evolve their reasoning dynamically, just as humans adjust their opinions during debate.
WP4 – Argumentative Reasoning
At the core of GRAIL, this WP integrates symbolic and neural AI to perform transparent, argument-based reasoning. It enables agents to build and update formal argumentation graphs (LPP), automatically generate explainable reasoning code (GORGIAS), and interact naturally with humans through arguments and explanations.
WP5 – System Integration and APIs
WP5 connects all project components into a unified multi-layer GRAIL architecture. It develops the necessary data structures, APIs, and interfaces to ensure interoperability between modules, enabling full system integration and paving the way for real-world testing and evaluation.
