|
The goal of the Machine Reading Program (MRP) is to create an automated reading system that makes the knowledge in NL texts accessible to any of an open-ended range of formal reasoning systems. FAUST will implement innovative solutions to the key challenges that arise in building this bridge.
We propose an architecture based on statistical joint inference over probabilistic relational models.
Supporting the simultaneous consideration of multiple random variables enables leveraging all mutually constraining information and integrating information across multiple sentences and texts. This
joint-inference engine will integrate information from multiple, more specialized inference modules, in particular, from an ensemble of natural-language modules. Such integration requires coordinating or
aligning representations from both multiple levels of analysis and multiple texts. We propose to continually improve these alignments by using machine-learning techniques. This approach, building on
significant recent advances in probabilistic representation and joint inference, will enable the system to consider the widest range of both linguistic and extra-linguistic evidence, using the same mechanisms that are used for the integration of linguistic information across multiple levels. | |
