NTL

Principal Investigator(s): 
Jerry Feldman

The NTL (Neural Theory of Language) project of the AI Group works in collaboration with other units on the UC Berkeley campus and elsewhere. It combines basic research in several disciplines with applications to natural language understanding systems. Basic efforts include studies in the computational, linguistic, neurobiological, and cognitive bases for language and thought. This research continues to yield a variety of theoretical and practical findings.

The NTL project is an interdisciplinary research effort to answer the question: How does the brain compute the mind?

More specific research questions include:

  • How can the brain -- a highly structured network of neurons -- support thought and language? How do the specific neural structures of the human brain shape the nature of thought and language?
  • How are language and thought related to other neural systems, including perception, motor control, and social cognition?
  • What are the computational properties of neural systems?
  • What are the applications of neural computing and semantic language processing?

Embodied Construction Grammar (ECG)

Embodied Construction Grammar (ECG) builds on decades of work on the Neural Theory of Language (NTL) project and other research on Construction Grammar and Cognitive Linguistics at ICSI, UC Berkeley and other institutions. It is the foundation for the tools and systems described on this page. A good general introduction is:
Feldman,J., E. Dodge, and J. Bryant. "A Neural Theory of Language and Embodied Construction Grammar"  In  B. Heine and H. Narrog, eds Oxford Handbook of Linguistic Analysis OUP, 111-138.

ECG is designed to support the following functions:

  • A formalism for capturing the shared grammar and beliefs of a language community.
  • A precise notation for technical linguistic work
  • An implemented specification for grammar testing
  • A front end for applications involving deep semantics (below)
  • A high level description for neural and behavioral experiments.
  • A basis for theories and models of language learning.

An application has been developed by Luca Gilardi to view and develop ECG grammars, as well as analyze input text:

A tutorial can be found here. There are also sample grammars that can be downloaded for experimentation:

Designing Systems for Natural Language Understanding

Natural Language Understanding (NLU) was one of the main original goals of artificial intelligence and cognitive science. We have implemented a system that supports full NLU for tasks of moderate complexity. Our natural language interface is based on Embodied Construction Grammar and simulation semantics; our primary implementation supports human dialog with an agent controlling a simulated robot, but is flexible with respect to both input language and output task.

natural language understanding

A version of this system developed for the domain of robotics can be found on GitHub, and downloaded for experimentation here. The instructions for running the demo can be found here.

A more in-depth explanation of the system was published in the proceedings of the 2015 Human-Robot-Interaction conference, and be found here.