STUDENT (SimulaTion and UnDerstanding of EveNts in Text)

ABSTRACT: Although Natural Language Processing research has produced powerful techniques for parsing and data extraction, computers remain mostly oblivious to the meaning of the language they process. Computers cannot, in general, connect the words and phrases in language to a larger model of the world that permits reasoning about the implications of what is written or said. We propose an algorithm that analyzes text-based language data using a method of inference designed to match the way humans process and describe activities and events. Our approach to language understanding combines text with external knowledge encoded in a flexible and expressive structure called an X-net. X-nets, invented by DAC team member Dr. Srinivas Narayanan, act as abstract and computationally efficient simulation of activities, states, and events. Unlike other inference techniques, X-nets make it practical to perform inference on language describing complex, uncertain, and interrelated events that unfold over time. We will evaluate our text inference capability using the same evaluation measures used to assess reading comprehension in middle and high school. BENEFIT: If we are successful, the technology developed under this effort will represent a major step towards the development of algorithms that achieve human-like understanding of text. Computers will be able to find and react to language data based on its meaning and implications, rather than the surface form of the words used. Existing stores of language data will become enormously more valuable once we can extract information based on implications instead of key words. And general-purpose, meaning-aware language understanding algorithms will act as the foundation of a new generation of data analysis and human interaction tools. In Phase I of this project, we focus on understanding the language of the limited domain of disasters and disaster response. Humanitarian Aid / Disaster Relief (HADR) is important to many Government, NGO, and private organizations. The results of Phase I of this project will be immediately applicable to a number of our current and potential customers who need exploit text data generated from large-scale, rapidly evolving events such as natural disasters.