Featured Research: Neural Theory of Language

The Neural Theory of Language is a comprehensive theory that explores how the human mind learns, understands, and uses language to communicate. It uses computational models and simulations of language and learning to answer basic questions about the production and use of natural language. For the past two decades, ICSI researchers have studied this relationship between the mind and language.

NTL theory research at ICSI is focused on answering the following questions:

  1. How can the brain support thought and language? How do the neural structures of the brain shape the nature of thought and language?
  2. How are language and thought related to other neural systems, including perception, motor control, and social cognition?
  3. What are the computational properties of neural systems?
  4. What are the applications of neural computing?

The thesis projects of three ICSI students, Nancy Chang, Eva Mok, and John Bryant, attempt to answer parts of the first two questions. (see page 4 for details). Lisa Aziz-Zadeh, a former ICSI post doc now at University of Southern California, uses fMRI technology to track what physically happens in the brain while it processes language (see ICSI Gazette, Vol. 4 No. 1, September 2005 for more on Aziz-Zadeh's fMRI experiements). Her fMRI experiments provide physical evidence in support of NTL theories about the relationship between the neural structures in the brain and language.

NTL answers many questions about the brain and language, and through basic research in several disciplines such as computer science, linguistics, neurobiology, and cognitive studies, provides a basis for practical applications to natural language processing systems. While theoretical NTL research continues, a group of ICSI researchers, led by current AI group leader Professor Srinivas Narayanan, are developing some of these practical applications based on NTL.

Question answering technology is one such application. ARDA's AQUAINT program, which ICSI has been involved in since it started a few years ago, enters Phase III this fall. Narayanan and his team at ICSI will be working closely with colleagues at the University of Texas at Dallas during Phase III. The ICSI team, through all phases of AQUAINT, has made use of NTL principles, particularly event modeling, in the development of intelligent question answering technology for computers. Event modeling can improve question answering technology by providing an intelligent template that describes a situation or event, providing keywords and background information that the software can use to search for potential matches in a set of data. Deep inferencing techniques and corpus based techniques are used for deriving the conceptual semantics needed for question answering systems.

A related application of NTL is semantic extraction, the use of semantics to access information. Many of the same techniques used in question answering can be applied to semantic extraction. ICSI is working on two semantic extraction applications, one for CISCO and one for Ask (formerly known as Ask Jeeves). The model of actions, processes, and events developed within the NTL project provides a natural, distributed operational semantics that may be used for simulation, validation, verification, automated composition, and semantic extraction.